DE4132754A1 - Arrangement for testing objects for surface defects - has transport unit, illumination, sensing, positioning, control, evaluation and ejection units for automated testing - Google Patents

Abstract

The arrangement for testing products for surface defects consists of a transport unit (1), an illumination unit (4), a positioning unit, a sensing unit (2), a control unit, an evaluation unit (7) and an ejection unit (6). The transport unit (1) carries the objects into a test zone in the correct orientation. The objects are illuminated (4) with light of defined different wavelengths either in a timed sequence and/or simultaneously. The light intensity for each image point and each wavelength region is measured and stored. The evaluation unit (7) compares the detected values with reference values and makes a good/bad decision for use by the ejector unit (6). USE/ADVANTAGE - Esp. for use in testing mine pins for colour, texture and stamping defects. The arrangement enables automated, successive and reliable testing of charges of different types of pin.

Description

The invention relates to an automatically operating device, all types of pencils or similar objects, that are flawed on the surface, recognizes and automatically sorted out. This device is said to either directly coupled to the manufacturing process, or they is an independently working device, the pins should be automatically removed from a pen memory. Several pens / second (e.g. 5-10) should be processed can be.

Mine pencils consist of a wooden casing, in the In the middle of a mine is inserted, the z. B. from graphite or some other dye-bearing material. The Pen variety is extraordinarily large. The pen shapes are largely standardized. They are common round or hexagonal and can be of different lengths be. The pencils are usually varnished several times and usually one or two colors, with either each adjacent Side of the hexagon pins bears different colors or only the edges are painted in a different color. About that can also be found on each pen for marking the Manufacturer, the hardness and color, etc. or as advertising text at least on one side at a defined position Embossing. Furthermore, multi-colored printed pens (offset printing) in trade. The pins can be unpainted at the other end, painted in different colors or with an eraser be provided.

The following errors are possible, for example:

Paint blobs
Missing color
Blurred color
Crumbled paint
Foreign bodies on the surface
Faulty or incomplete painting of the end of the pen
Color blurred or missing on pen edge
Wells in the wood
Grooves in the wood (e.g. at the glue point)
Broken wood
Missing imprint
Embossing in the wrong place
Blurred embossing
Embossing not in the middle
Embossing damaged wood
Unclean pressure

The final inspection is currently carried out by hand. With machine The pens can aid in the visual field of the controller. It can still be Make sure that the pins roll off so that the controller sees every side of a pen. The real one however, controlling and controlling activities are the responsibility of the human observer. With such technical aids can control speed of several Reach pins / seconds.

An obvious generalization is to the Place the human eye to put a camera, where the image evaluation and control of the process now from one Data processing system is made. In addition to error detection then the computing unit also has the position detection of the pens and - in the case of moving pens - the To carry out pen tracking.

A reduction in the computing effort can be achieved through Stiftpo Positioning and synchronization of the pen movement with the Reach camera image capture. The test object with a gripping device in a predetermined position recorded and then triggered the test process.  There is no need for position detection and pin tracking.

1

As an introduction and to introduce the terminology used basic knowledge of the Classification methods reproduced. It is shown how a pattern is received, how it is pre-processed, what features can be extracted and how based on these characteristics, a classification is carried out can be. Finally, learning classifiers discussed. Extensive explanations are in the Classification of patterns, Springer-Verlag, Berlin and Pattern Analysis, Springer-Verlag, Berlin by H. Niemann to find.

In the present case it is a so-called classic problem, where there are two classes. The first class contains all flawless and the other takes all faulty Pens on. On the other hand, should a more detailed investigation to be made z. B. an unmistakable one Description of the appearance of the pens will be delivered so there is an analysis problem. Nevertheless, one can Classification problem include an analysis problem if e.g. B. a medical diagnosis is made, the result is "sick" or "healthy". A classification or In principle, the analysis problem relates to a restricted one Problem area.

The representations to be examined are called samples. A pattern can e.g. B. an image, a sequence of Pictures, an acoustic or any other electronic Be a signal. A classification or analysis of a complex patterns, such as B. the image of a pencil, is only possible if this pattern into individual, simpler Components can be disassembled, this being certain  Have characteristics that belong to the components characterize to a certain class. The Classification is generally done using numerical methods. Each pattern becomes a set of features extracted, which form the so-called feature vector. Essentials A prerequisite for a successful classification is that the characteristics of the pattern of a class differ from the characteristics the patterns of the remaining classes differ significantly. For example, has a pattern that is exactly one in four Classes can be assigned two one-dimensional characteristics and is the space spanned by these features two-dimensional, then e.g. B. the individual classes in Cartesian feature space can be determined by the fact that the Characteristics of each class lie in a different quadrant.

In general, there is a pattern in the present Work is an image, from a variety of samples. So owns a color television picture no less than 512 × 512 pixels for each of the three color channels red, green and blue. This corresponds to 786,432 intensity values. Everyone Value can be seen as a characteristic. However, since this are mostly highly correlated, is calculated from computing time and Data compression performed due to storage space. So for example B. to differentiate between round and rectangular Objects the ratio of the length of the outline in the square to the area of the object. Such a feature is called Designated form factor.

The real classification problem is due to the location of the features of a pattern in the feature space determine the class belonging to this characteristic or model. An analysis problem can e.g. B. be solved by by first dividing the pattern into simple components - so simple pattern - is divided to this one classi application process. In the end, all classes become sification results combined, from which a Description of the pattern or an overall classification  result follows.

A classification problem is expediently included in the four process steps

Recording the pattern,
Preprocessing,
Feature extraction and
classification

subdivided. In addition, a learning step is required to close the area in the feature space from a sample determine in which all possible characteristics of patterns in a class.

The object to be examined is used to record a pattern scanned at as many locations as possible, d. H. discretized. The intensities are usually for each pixel of the three color channels red, green and blue. The Image quality and also the sampling time depends on the used Recording technology. So process z. B. TV cameras 25 frames / second.

Discretion creates a loss of information, so that the recording is not the exact replica of reality reflects. Additionally, through systematic or accidental errors of the recording device falsified the image. So z. B. when shooting moving objects or a moving recording device geometric distortions if the individual image lines are in order be included.

The goal of preprocessing is to use the ones mentioned above Undo faults as much as possible and the Process samples in such a way that the following classification simplified and the classification result improved becomes. The preprocessing is essential for a successful and reliable classification. In spite of the importance of certain prior  practically not theoretically demonstrating work measures. The methods used here are therefore largely heuristic in nature, the use of which is sensitive to the developer subject to. This also applies to those below discussed feature selection.

Can distinguish objects by their contours alone the data volume can be reduced by a threshold operation significantly reduce. The result of this Preprocessing step is a black and white image with only two gray values.

Another way to reduce data is to the samples of a data compressing transformation subject. So z. B. the same speech signal in several long intervals are divided, with each interval a frequency analysis is made and the amplitudes of the individual frequencies can be saved. These The method can also be used in image processing. The The Fourier transformation provides mathematical tools for this ready.

Other preprocessing measures are used to reduce the Variability of parameters of the pattern. For this, parameters, that are not relevant to the classification standardized a constant size. So z. B. characters standardized to a certain size and width and any Corrected skew. You can also change their font size normalize.

Still other measures are meant to discretize the quality Improve patterns. For example, a noisy one Patterns can be enhanced by smoothing by using the Gray values of neighboring pixels by means of average values over several pixels from this area are replaced. By The amplification of the shares of high frequencies can be Highlight contours, which is at least a visual improvement  which means sharpness. This is especially true with nebulized Recordings of importance. By means of Histogram evaluations can make corrections Overexposure or underexposure can be compensated. More demanding mathematical methods, the so-called filtering are assigned, are used to counteract the disruptive influences from clouds when taking satellite pictures. Furthermore - as already mentioned - geometric Corrections are made.

The characteristics of patterns of the problem area must be selected in this way be that the patterns of different classes distinguish as best as possible based on their characteristics and features of patterns in a class differ as little as possible differentiate from each other. Characteristics are desirable the invariant to translation, rotation and resizing are.

The simplest features are the samples themselves. These generally contain a lot of superfluous information, the entrained for the further process steps and must also be edited. For this reason - and not least because of the immense amount of data - divide in the in most cases the samples as characteristics.

Features or feature vectors that are relatively easy to calculate on a heuristic basis arise from linear Transformation of the samples, the value matrix of the Patterns are developed according to orthonormal basis vectors. This creates uncontrolled features. example One such development is the Fourier transform. The n largest Fourier coefficients are used as characteristics used.

Other features, the translation, rotation and sizes can be invariant, can be different by means of the moments Determine order or combinations thereof.  

The autocorrelation coefficient of the is also often increased examining pattern or a part thereof calculated a known pattern and used as a characteristic. Key figures are characteristics that are based on purely intuitive and expe rimental basis. This includes, for example the form factor mentioned above or z. B. saving the Pattern only at certain points or lines, the projection of the pattern onto a dimension or also the Specification of the gray value histogram of the pattern.

In contrast, analytical methods first calculate for one Class of transformations the transformation matrix so that for a classified sample a certain one Quality criterion is optimized. This transformation is on applied each of the patterns to be examined. The trans The result of the formation is the desired feature vector. Important Examples of such quality criteria are the inter and the intra-class distance. The inter-class distance gives that mean square deviation of the characteristics of a class from the characteristics of another class, while the Intra-class distance to the mean square distance that relates characteristics within a class. With good ones The inter-class distance should be maximum and the Intra-class spacing must be minimal.

As already mentioned, the classification corresponds to one Mapping of the characteristic space to the set of all classes, the feature space is partitioned such that in the normal case every feature or feature vector is clearly accurate a class is assigned.

The occurrence of observed patterns or features can to be viewed as a process of statistical laws is subject. Are the class-related for each class Distribution densities are known, so a statistical classification procedure be applied. Is a classified Sample available, it can be  statistical tests the type of underlying distribution density determine. With processes such as the Maximum likelihood method, then its parameters estimate. If the density cannot be specified analytically, then it can be specified at least in the form of a histogram. To Classification makes every possible class a test variable calculated. You decide on the class, its test variable takes an extremum. As a test variable z. B. the A- posteriori probability can be used. The decision is thus in favor of the class, for which is the probability of this class on the condition the appearance of the observed feature vector among all other probabilities is greatest. Such one Classifier minimizes the likelihood of a wrong decision to meet and is used as a Bayesian classifier designated. The test variables of this classifier can be simply calculate if the feature vectors are class by class normally distributed and in addition the components of the feature vectors are stochastically independent.

The statistical method mentioned has the disadvantage that the conditional densities of each class must be known. Only the normal distribution is currently mathematical treated. Distribution-free classifiers use as Decision function also test variables, but of are independent of the distribution. For this, the components a transformation matrix so that the classification error for a pre-classified sample becomes minimal. For example, the expected value of the quadratic distance of the decision function used of the value of the exact decision function for one pre-classified sample determined. The exact decision function for an n-class problem is an n-dimensional one Vector that is in the place of the class that is the feature vector is assigned, e.g. B. a 1 and all others Positions has a 0. The solution arises with the help the regression analysis.  

An example of a parameter-free classifier is the next one. Neighbor classifier. The characteristics of a classified sample together with the corresponding Class saved. The characteristic of an observed pattern becomes the characteristic from the stored quantity of characteristics selected that is closest to this characteristic in the characteristic space lies. The pattern now becomes the class of its neighboring feature assigned.

As can be seen from the above, the existence is a basic requirement of a classification. The parameters of some classifiers are not predictable without this sample while this is for others The procedure is only given implicitly, namely if certain Parameters based on experience and heuristic considerations of the developer. The sample should also be classified, e.g. B. done by hand can.

Usually, using the classified sample the classifier is dimensioned at the beginning. This takes place in the so-called learning or training phase. The Classification parameters then remain unchanged.

Other classifiers used as learning classifiers are able to improve their classification performance based on the observed patterns or by constant Switch between learning and classification phase too improve. However, the learning only relates to the improvement or adjustment of the values of the classification parameters, but not on the type of the classifier or the extraction of features. There is no general one yet Method to learn the features too improve.

If the sample is classified, it is monitored by one learning classifier spoken. The learning process with  Patterns whose class is not known is considered unsupervised Learning means.

2nd

The following are those used for error detection in here Problems occurring in this case are discussed and it basic solutions are shown.

Due to the variety of types of the pens and for simplification the handling of the system should be able to automatically adjust to new pen types. It is a Provide mechanism for knowledge acquisition. This enforces a two-step system approach at Checking pens of a pen type: First, the System in a training phase a sufficiently large number passed by error-free pens of this type. From this The system determines all learning samples for classification of "good" pens necessary characteristic values and statistical Parameters. Furthermore, self-tests can also be carried out in this phase as well as a calibration of the system. A Quality improvement can be achieved if the system is additional also receives a negative sample. After finishing the Characteristic extraction is the system in the following Classification phase - which is the actual work phase represents - able to produce the pens of this type with regard to the quality of their surface and to discard the pens identified as defective.

Closely related to the knowledge acquisition at the start of the system is the ability to continuously improve the classification parameters for new pens. This is required because the parameters are subject to fluctuations in production (e.g. slight change in color), aging of the Components (e.g. decrease in the intensity of the lighting), Temperature fluctuations etc. continuously, but only slowly  change. Common methods for dynamic adjustment of the Characteristic values are e.g. B. moving average and variance calculations, where only the values of the Pens recognized as "good" and the last ones received Measured values are weighted higher than older measured values. There such operations are usually very time-intensive, it may be useful to have one or only a few for each pen Adjust parameters.

The measurement method is based on the pin on as many as possible To scan points. The conventional television picture recording technology, based on light beam scanning is too imprecise. Also, a lot of the picture isn’t used, because the pens are an inconvenient compared to the image size Have a ratio of length to width. Others more precise Methods such as B. vibration mirror scanner, need one too long recording time. An alternative is a special one Semiconductor scanning unit, the as many recording elements as possible has so that parallel processing is possible is.

The only measurable parameters that also the human eye perceives, wavelength and intensity are visible Radiation. Together, these two parameters form the spectral characteristic, each color or mixture of colors clearly characterized. A fault is recognized that their spectral characteristic is different from the spectral characteristic the expected color significantly. An integral measurement, i. H. the measurement of the total intensity over the full spectral range or one larger part of the spectral range, as the usual Perform semiconductor devices, not very helpful. However the determination of the entire spectral characteristic is too time-consuming and also not necessary. Instead, it is tolerable Loss of information possible for every pixel make selective measurements as measurements of intensities for multiple wavelengths. The missing parts are then  due to the change in intensity with respect to these wavelengths recognized. In order to discretize the spectral characteristic loss of information arising as little as possible hold, the wavelengths must be determined for each color, the most related to their classification, d. H. Distinction to other colors.

So it will be necessary for each color to be examined to make another selection of wavelengths, too the number of wavelengths to be measured fluctuates. These The reasoning is supported by the fact that in the case of a There are usually no new mistakes at this point Color is added, just the one at this point expected color replaced or covered by another color is in which the pen is painted and with it is known. In other words, it can be assumed that Colors of imperfections are also pen colors. So is too expect that not any spectral characteristic as Spectral characteristic of a defect can occur, but only a subset of it, namely combinations of spotting tracing characteristics of the pen by varying the color intensities. As an additional color, however, is the color of the untreated wood.

The usual color television camera technology with the three spectral channels red, green and blue is for color differentiation and not sufficient for reliable classification. A multi-channel scanning device is therefore used for image recording used whose spectral channels are targeted set the color spectrum of the pen to be examined to let.

Unlike the conventional classification methods, where in the learning phase only the values for classification required parameters are determined here too the parameters, namely the wavelengths themselves, are also determined.  

Some pinholes are depressions, whereby whose color is flawless. Let this deepen do not demonstrate shadow formation that is homogeneous and diffuse illumination, at least for smaller bumps is not present, these errors can only be determined using the Change in light intensity due to the greater distance to the recording device can be recognized. Because the intensity with the square of the distance from the cradle decreases, it should be as close as possible to the surface of the pen are located.

Due to the high throughput rate, the computing time is Keep the classification as low as possible. So that's it basically necessary, the computing unit by constructive Relieve measures as much as possible. The Use of extremely computing-intensive algorithms such as B. finding a pencil (contour finding), moving detection and tracking of a pen as well as standardization and corrective measures (e.g. shadow removal, Layer normalization, homogenization of inhomogeneous illumination) can by appropriate mechanical, optical and electronic Aids are avoided. So it is necessary the pins using a mechanism as precise as possible to position the image acquisition using the Synchronize pen movement or positioning and the illumination as homogeneous and shadow-free as possible shape.

In the simplest case, the pins can be used during the sampling time thereby being held in a certain position by guided by one or more movable stops or be prevented from moving. After done The strokes reflect the held pen free. Another option is to use the pens Lock with the help of a gripping device. To avoid The position accuracy of position standardization measures smaller than the dimensional tolerances of the test objects or  the manufacturing tolerances of the paint to be tested. For position detection, it may also be necessary in the Learning phase to determine a positioning feature. As a positioning feature can e.g. B. a clear and easily identifiable Serve embossing, which is with each pen on the must be in the same place, or a dedicated one Purpose attached to each pin in the same place Color marking.

Optical sorting devices exist e.g. B. for bottles car washes to get damaged or after washing discard dirty bottles. Be too with machine help e.g. B. discarded stained potato chips. The devices used here can however take advantage of the fact that there are very few types - or just a single type - which is also simpler Are structure so that the device does not have the ability to owns need to add new ones by itself Types.

In DE 27 46 615 and DE 29 24 605 there is a manual selection of working wavelengths that provide an optimal color distinction enable considered. DE 27 46 615 serves the classification of objects with homogeneous color distribution, whereby the color distinction is limited to two colors. DE 29 24 605 is used to differentiate colors in Framework of the authenticity check of banknotes, but can not recognizable due to its integral working method of the defects here are used. From the EP 00 65 363 is a scanning unit for color differentiation known, the light striking the scanning element with Using mechanically moved color filters in its spectral Shares can be broken down. The scanning unit is suitable for manually switching filters, but allowed not for optical quality control of mass products required rapid switching of spectral ranges. The position of the filter holder in front of the scanning cell  proves to be disadvantageous because of the movable filter carrier entrained dust the optical properties the scanning unit deteriorates. Second, there are constructive ones Difficulties due to the space requirements of the Filter carrier.

The invention has for its object the device automate the type mentioned above, so that they avoid the disadvantages of the prior art Technology as well as taking advantage of special properties of Test objects batches of any type of pens or The like can reliably check one after the other.

This problem is solved with the help of the characteristics of Claim 1. Advantageous further developments of the object of the claim result from the subclaims.

3rd

The invention is based on in the drawing illustrated embodiments further described.

Fig. 1 shows the basic circuit diagram of the device.

Fig. 2 shows a filter carrier with 12 color filters.

Fig. 3 shows a filter support for a graduated filter.

Fig. 4 shows a lighting unit with two physical channels.

Fig. 5 shows examples of various embodiments of Abtastmatrizen.

Fig. 6 shows an example of an embodiment when scanning with the aid of optical fibers.

FIG. 7 shows an example of an embodiment in the case of scanning with the aid of lens optics.

Fig. 8 shows the pin movement in a scanning tube.

Fig. 9 shows a scanning device divided in two halves.

Fig. 10 shows the basic structure of the positioning device for hexagonal pins.

Fig. 11 shows the development of a hexagonal pin.

FIG. 12 shows a possible region division of the pen from FIG. 11.

FIG. 13 shows a possible data structure for the region classification of FIG. 12.

FIG. 13a shows the geometry for data block structure of FIG. 13,

Fig. 14 shows the block diagram of the learning phase.

Fig. 15 shows a plane spanned by three corner colors and spectral color mixtures thereof.

FIG. 16 shows the convex hull of the spectral vectors of FIG. 15.

FIG. 17 shows a smoothed convex hull and calculated spectral vectors of the corner colors for FIGS . 15 and 16.

Fig. 18 shows the functional structure of the system.

Fig. 19 shows the block diagram of the preprocessing.

Fig. 20 shows the block diagram of the classification unit.

Fig. 21 shows the structure of the unit for the classification of homogeneous regions.

The lighting unit, the scanning unit, the positioning unit and the computing unit explained. An overall description of the Contraption. The description is limited to pencils, although the invention is similar to others all kinds of objects is applicable.

The entire device, which is shown in Fig. 1, consists of a transport unit 1 , which receives the pins and transported in a suitable manner. In Fig. 1, the pins are moved from right to left. The drawing device, which is necessary to remove the pens from a pen memory, which is also not shown, or to take them over from the production machines, is not shown. The transport unit is intended to transmit the pens in an orderly and separated manner, so that only one pen or a predetermined number of pens can be transferred to the scanning unit 2 and all pen tips point in a predetermined direction.

The transport and the separation of the pins can, for. B. by pushing rods or the like and / or conveyor belt or one Transport chain can be made, the belt or the Chain has bulges for receiving one pin each, so that the parallel pins are in a perpendicular to their longitudinal axis moving direction. Here can be assumed that all from the pen memory or taken over from the upstream production machines Pins have the same orientation.

The scanning unit 2 is used to optically recognize the pen surface, the pins being brought into a position suitable for scanning if necessary by means of a positioning unit 3 . In the simplest case, this can be accomplished by a locking mechanism, the test object being pressed against stops in order to achieve a vibration-free fixation of all test objects in a defined position. If, on the other hand, fine adjustment is required, the positioning unit also implements a defined displacement of the test object in the longitudinal direction and the initiation of a specific rotational position.

The lighting unit 4 supplies the necessary radiation, which is fed to the scanning unit by means of an optical waveguide 5 .

An ejection unit 6 is located behind the scanning unit 2 in the transport direction. The ejection unit separates the pins recognized as rejects. All the remaining pens, which are thus recognized as good, are forwarded by the transport unit to a collecting device, which is not shown in FIG. 1.

The overall sequence and all partial sequences are controlled with a data processing system 7 , which send signals to all units and, if appropriate, can also receive signals from them. To monitor the process and to detect irregularities, it is necessary to use light barriers or other signal devices.

4th

The lighting unit 4 supplies the radiation required for the image recording. It is of fundamental importance for image processing that the pens to be examined are illuminated as homogeneously and stably as possible. In order to avoid interference from extraneous light, the scanning unit is lightproof. The radiation generated by the lighting unit is supplied by means of an optical waveguide provided for radiation monitoring or for lighting purposes.

The pens to be examined are at as many points as possible to feel. It should be precise to everyone Spectral characteristic belonging to the sampling point can be determined. There this is too time consuming, you will limit yourself to the intensity only for special wavelengths or special ones To determine wavelength ranges. The number of to be measured Intensities as well as the wavelengths or wavelength ranges itself depends on the colors of the pens and how difficult it is to compare the colors with each other and also from To reliably distinguish defects. The wavelengths  are therefore for each type of pen taking this into account Redefine conditions.

These wavelengths are called measuring wavelengths below. a measuring wavelength also a larger waves longitudinal area can represent. In practice describes the Measuring wavelength always a spectral range or a spectral Sensitivity range, the intensity or the spectral sensitivity at the measuring wavelength Owns maximum.

It turns out to be the color of a monochrome test object green as advantageous, with a wavelength (or a narrowband wavelength range) a and two broadband Wavelength ranges b and c to work. The wavelength a lies in the green spectral range, being the maximum the radiation intensity (i.e. the measurement wavelength) with the maximum of the spectral characteristic of the color used should match. The wavelength ranges b and c should the range of visible light to the exclusion of the Cover the spectral range of the green color used. So b encompasses the red and yellow spectral range, while c encompasses the blue and violet spectral range covered.

For a hexagonal pen, the six sides alternately painted in one of the colors yellow and blue the five wavelength ranges turn red, yellow, green, blue and violet are particularly advantageous. The Measuring wavelengths for yellow and blue are expediently in the intensity maxima of the colors used and their Wavelength ranges are more narrow-band. The yellow areas with the three spectral ranges red, yellow and green + blue + purple checked while under investigation of the blue areas the three spectral areas red + yellow + green, blue, violet can be applied.

Are many colors or for painting an object Colors with close spectral ranges used, it may be helpful for a better distinction  by color additions the color differentiation also on the area the invisible light, e.g. B. the ultraviolet Area to expand. So could be with two similar colors a color can be mixed with a color additive that is in the Range of the ultraviolet radiation its maximum intensity have.

One way to measure the spectral characteristic only at a certain point To measure the place of the frequency spectrum, there is therein, the measuring point with a radiation source z. B. a Standard light can be to illuminate and by means of a filter only the frequency parts to be measured to drop an integrally measuring scanning cell. This The principle is based on the usual spectrophotometers. The Standard light is characterized by the fact that it is light of all Emits wavelengths, the spectral distribution of the Light is known.

An alternative is from the radiation source only the radiation of the desired wavelength or To have the wavelength range emitted. This eliminates the subsequent filtering. This is for solving the This task is important because the scanning cells should be very close to the mine pins and therefore only a few constructive options are available stand to introduce a filter device. It is coming complicating the fact that a large number of scan cells are required and only one filter attached there is automatic must be changed.

An obvious way to generate colored light consists in appropriately illuminating a light source filter. The lighting source is an optical one Subsequent filter disc, its wavelength also should be automatically changeable.

In FIG. 2, a disk-shaped filter holder 8 is introduced, the filter discs 12 can accommodate 9. The radiation from the radiation source, for example, only falls on the filter disk located at the top. Changing the filter disc means rotating the filter carrier onto another filter disc. The rotation can be done with the help of a motor such. B. a stepper motor can be realized. The marking 10 serves to identify the first filter disk. So that the filter carrier z. B. with a light barrier - if the marking 10 is a hole or a pin - set to a defined start filter. If necessary, a similar positioning mark should be made for each filter. Furthermore, a hole provided for receiving the filter disks could be empty, which is designated by the color "white" in FIG. 2, while there should also be the possibility of blocking the radiation for one position. This position is marked "black" in FIG. 2. Both positions can e.g. B. used for calibration purposes.

The filter carrier variant proposed in FIG. 2 can only accommodate a limited number of filter disks. This restriction can be avoided if, instead of individual filter disks, which are each intended for a wavelength or a wavelength range, an optical gradient filter is embedded in the filter carrier 11 , as shown in FIG. 3. There is again marking 10 and positions 12 and 13 for "white" and "black". 14 shows the gradient filter. In this way, filtering can practically be carried out for every possible position in the frequency band of the gradient filter 10 . A prerequisite for this is, however, that the radiation hits the filter disk as punctiformly as possible or that light is mixed with the aid of optical methods after the filter disk so that radiation with the same wavelength as possible occurs at all points of the pen.

The principle of the filter carrier of FIG. 3 for a circular gradient filter can also be used for a straight gradient filter. In this case, the desired filter positions are set with a linear movement. This can be done e.g. B. return to a circular movement by means of a drive spindle. A longer positioning time must usually be accepted.

In FIG. 4, a lighting unit shown has two physical channels. It consists of a radiation source 15 . This could e.g. B. be a halogen lamp. The type of lamp and its parameters (color temperature, intensity, etc.) are to be determined by tests and measurements. A unit 16 , which is used to process the radiation, connects to the radiation source 15 . This can e.g. B. be a common capacitor, which also bundles and preprocesses the radiation in such a way that it can be fed as optimally as possible into an optical waveguide 17 . This splits into two paths, each of which leads to a filter carrier 18 and 19 , which can be constructed according to FIGS. 2 or 3. The filter carrier 18 and 19 are, for. B. with a motor 20 and 21 , which, for. B. stepper motors can be driven. Downstream of the two filter carriers is an aperture 23, also driven by a further motor 22 . This should switch from one optical waveguide channel 24 to the other optical waveguide channel 25 , block the radiation or also let the radiation from both channels pass. If the diaphragm 23 has the possibility of interrupting the radiation path, the "black", ie the translucent, region shown in FIGS. 2 and 3 is unnecessary.

The additional aperture 23 is therefore introduced in order to ensure the fastest possible switching from one channel to the other channel. Switching would also be possible directly with the filter carriers 18 and 19 , but the opaque area would then be necessary. However, the inertia force that occurs due to the mass of the filter disk and the expansion of the filter carrier during switching leads to a long switching time, while the diaphragm 23 can be designed with a significantly lower mass and also smaller dimensions.

A mechanically driven diaphragm is shown in FIG. 4. It can also be an electro-optically operating radiation barrier z. B. on the basis of liquid keitskrallen. There is a liquid crystal plate in each radiation path, the light transmittance of which is controlled by a common control unit. Materials that regulate the radiation flow of polarized light depending on the current, voltage or magnetic field are correspondingly suitable. An electromechanically working diaphragm with the aid of piezoceramics would also be conceivable.

The radiation path adjoining the diaphragm 23 consists of one optical waveguide 24 and 25 for each of the two physical channels. The two radiation paths are fused in a switch 24 to form a single radiation path with a resulting optical waveguide 27 .

When using a graduated filter, the radiation on Filters are fed in and forwarded in such a way that possible little radiation with different wavelengths is produced. The fibers of the optical fibers in question should therefore be arranged so that the area constant wavelength of the gradient filter as good as possible exploit. The individual fibers should be radial to The center of the filter holder.

With the principle shown in FIG. 4, lighting units with three or more physical channels can also be produced. This means that all measuring wavelengths of a pencil type can be permanently set in the learning phase. There is no need to switch the filter carrier in the classification phase, which relieves the mechanical workload and, if necessary, speeds up the classification process by eliminating the switchover time. The lighting unit is light and dustproof.

As expected, the radiation source is in the range of visible light work, so that also from light source and one can speak of light instead of radiation. Furthermore should consider polarized light to reduce reflection to be pulled. The filtering also does not become light exactly one wavelength, but a wavelength range let through. The red painted pen of the example are broadband spectral ranges b and c desired. For checking the surface of the pen is first with the Measurement wavelength a is applied. The light of the measuring wavelengths b and c can either be a mixture of light at least one two-channel arrangement or sequentially generated. In the first case there is no test step. Checking with The use of light mixtures is particularly feasible if the sensitivity curve over the spectral range in question is largely constant, so that wavelength-dependent Standardization measures can be dispensed with. Analogous to the one-color lacquered pen, for example in the above-mentioned yellow-blue painted pen four Test steps are sufficient.

5

The scanning unit is described in more detail below. It a distinction is made between three variants.

In the first variant, the scanning unit consists of a Matrix of scan cells, with as many scan cells as possible should be present that a pen in one step can be scanned. Are integrated in the scanning matrix  Optical fibers, which have a largely homogeneous illumination of the entire pen surface. As scanning cells serve preferably light-sensitive semiconductor elements, such as B. CCD elements (Charge Coupled Devise), photodiodes or photo transistors, because of the smaller design primarily CCD elements or photodiodes come into question will.

FIG. 5 shows various designs of Abtastmatrizen for different types of scan. The elements 28 and 29 are the scanning cells, while the circles 30 represent fibers or bundles of fibers of optical fibers. The size ratio of the lighting elements 30 to the scanning cells 28 or cannot be seen to scale. The number of fibers belonging to a lighting element 30 and the type and diameter of the fibers in connection with the lighting unit are to be dimensioned such that the illuminance on the pen surface is sufficiently large.

The individual fibers of the lighting elements are combined to form an optical waveguide. This is marked with position number 5 in FIG. 1. The optical waveguide 27 shown in FIG. 4 in turn corresponds to this optical waveguide.

The scanning device of the scanning unit consists of scanning matrices. The scanning matrix could, for example, be produced in such a way that the lighting elements 30 are cast or glued together with the scanning cells 28 and 29, respectively.

The second variant for forming the scanning device consists in not incorporating the scanning cells 28 and 29 into the scanning matrix or scanning line. The schematic drawing of an embodiment of this variant is shown in FIG. 6 using a scan line. The scanning line consists of optical fibers 38 for illuminating the surface of the test object. These optical waveguides 38 can also be bundles of optical waveguides in order to generate light with sufficient intensity. Each optical waveguide 39 is connected to a scanning cell 40 in such a way that the light flowing out of the optical waveguide 39 radiates into the light-sensitive surface of the scanning cell 40 .

Ideally, the scanning cells measure exactly one point. In fact, the reflected radiation from this point as well as a small environment around it. This area may be called measuring surface. With the usual scanning cells the scanning center is weighted higher than the edge of the measuring surface. The shape and size of the measuring surface depends on the Design of the scanning cells. The scan cells of the scan matrix should be so close that the pen surface is complete can be measured.

The following measures can help to improve the scanning quality: The surrounding of the scanning cell or the scanning optical waveguide, the lighting optical waveguide are arranged so that the light flowing out illuminates the measuring surface homogeneously and not vertically, but z. B. falls at an angle of 45 degrees on the surface of the test object and thus causes reflections. This can be achieved, on the one hand, by installing the optical waveguides used for lighting in a curved or oblique manner so that they emit the light into the measuring point. Another possibility is to form the exit surface of the optical waveguide at an angle, so that the emerging light falls into the center of the measuring surface of the relevant scanning cell. This also applies to the scanning matrix according to FIG. 5 as well as to the scanning units according to FIGS . 6 and 7. Another possibility is to design the surface of the scanning cells 28 and 29 or of the optical waveguide 39 in such a way that a lens effect is achieved and that Measuring area is reduced to a possible punctiform measuring area.

The optical waveguide 39 can be as fine as desired. For one scan line as shown in FIG. 6 is a CCD array can be used as light-sensitive semiconductor component. The optical waveguides 39 arranged in a row should be so fine that exactly one element of the CCD line is assigned to each optical waveguide.

The third variant, like the first variant, does not use optical fibers for scanning. Instead, optics are used, as with recording cameras. The optics represent the incident light on a CCD semiconductor matrix behind it. FIG. 7 shows an example of an embodiment for this scanning principle. The lighting is carried out with the aid of the optical waveguides or optical waveguide strands 41 . They are arranged at an angle so that they illuminate the measuring surface evenly and generate as little reflection as possible. The scanning element 42 contains the optics and the CCD matrix at its upper end. The optics are set so that they map the image of the surface of the test object as sharply as possible onto the CCD matrix. The CCD matrix can be square or circular. If a scanning element 42 covers an area with a diameter of z. B. from 10 mm, a 160 mm long pin would require at least 16 scanning elements 42 , which are arranged in a scanning row. For the sake of simplicity, the lighting strands 41 are likewise arranged in a row.

Like the scanning cell of FIGS. 6 and 7, the CCD matrix only needs to process gray values, but no color values. The colored light required for lighting is supplied by the lighting unit. The color of the light can be set and changed automatically in the lighting unit. Optionally, an optical filter in front of the sensing elements 42 or the serving of the illumination optical waveguide, z. B. for correcting the sensitivity characteristic of the scanning elements or for filtering out light reflections (polarization filter). Commercially available recording elements for black-and-white images can be used as the scanning element 42 , with a constant sensitivity over the required wavelength range being desirable.

With a processing or decoding unit downstream of the scanning unit according to FIG. 7, the measured values of the scanning elements 42 are processed and stored in such a way that the stored measured values represent a discretization of the surface of the test object measured by the scanning elements.

This can in accordance with the arrangement Fig. 7 characterized happen that the measured values of the CCD matrix elements of adjacent sensing elements 42, the same or overlapping Measuringsurface measured on the surface of the test object, for example. B. be summarized by weighting and arithmetic averaging to a measured value. The measured values to be transformed and the transformation parameters are determined with the aid of one or more test patterns structured in a grid pattern after assembly and fine adjustment of the scanning elements 42 . One of the test patterns can, for example, be constructed like a checkerboard, with one field corresponding to exactly one measuring surface. If the pattern supplied by the scanning elements is irregular and not a 1: 1 image of the test pattern, there is an overlap of measuring areas of scanning cells. The addresses of the relevant scan cells, including their correction values, are stored in a memory which can be accessed by the said processing unit, which - based on these values - e.g. B. by linear interpolation, the correction of the measured values and thus a geometric correction - ie equalization - of the scanning grid. If necessary, fine positioning of the scanning elements 42 of the scanning row is helpful to simplify the geometric correction of the recorded images. This calibration is only to be carried out in the event of maintenance, while the measured value correction must be carried out as a preprocessing measure after each scanning process. The result of the processing unit is thus a uniform and undistorted rasterization of the pen surface scanned by the scanning row according to FIG. 7.

Another data preparation measure includes the compensation of sensitivity differences between the individual scanning cells 40 of the CCD matrix elements and the compensation of aging-related differences in brightness of the lighting unit. With the help of monochrome test objects that the system z. B. before the start of the learning phase or before starting work, the scanning elements are calibrated. So z. B. a correction value can be stored for each scanning element and for each measuring wavelength. In the classification phase, the correction memory is accessed for each scanning cell or scanning element, the correction value belonging to the set test wavelength is determined and multiplied by the actual measured value. The sensitivity characteristic can be standardized in the same way.

The scanner should have as many scan cells as possible possess that all measuring points are scanned at the same time can be. If there were only one scan line, it would have to Line by line are scanned, for which the one to be scanned Pen continuously rotated using the positioning device should be what slow down the recording time accordingly would.

The address reader of the Bundespost (test run 1978) used a scan line with 512 photodiodes that are 60 mm long is. This corresponds to 8.5 points per mm. The reading area is scanned with 512 × 1024 pixels. The address reader can process 17 programs per second.  

Various embodiments are possible for scanning in one operation. Thus, the sensing matrix could in FIG. 5 are embedded in a tube, Fig. 8 shows. The scanning variants shown in FIGS . 6 and 7 can be used in an analogous manner. The pins are inserted into the scanning tube from one side by means of a mechanism, and the scanned pins are removed on the other side. A modification that also allows the upper end of the pen to be examined uses a scanning tube sealed on one side. If this termination, which carries measuring cells, is designed to be movable, the principle of FIG. 8 can still be used, on the other hand the pins on the same side must be removed. It should be noted that the pins at the tip can only be scanned imperfectly, since the pins must be held at least at the tip with a mechanical gripping device. This is not a significant problem. The presence of the pen tip could be checked with the aid of stylus needles. On the other hand, the pins could be moved using a mechanical slider and e.g. B. with compressed air support completely in the scanning tube and removed with the help of compressed air from the tube.

A further alternative consists in the design of the scanning tube in the form of two half-shells according to FIG. 9. The pin 32 to be scanned is held in place by means of the positioning device 31 and, if necessary, positioned accordingly before scanning. 33 shows the upper half and 34 the lower half of the scanner. The scanning apparatus of FIG. 9 is designed so that also the upper pin end, however, the pen tip can not be examined. Furthermore, Fig. 9 assumes that the pencils have not yet been sharpened at the time of scanning. By appropriate modification of the positioning device as well as the scanning device, it can be accomplished that pointed pencils can also be handled. Here, the fingers of the gripping device are redesigned so that they at least partially assume the shape of the pen tip when closed. The pen tip can be examined at least mechanically with the aid of stylus needles. The presence of the sharpened lead can be checked using a light barrier.

To scan the upper half of the pen, the pen 32 is moved upwards and accordingly to scan the lower half downwards. So two scanning steps are required. The other end of the pen is guided by a slider from below when scanning the upper half and from above by a slider when scanning the lower half. The device can also be designed so that the pin is stationary and the scanning halves 33 and 34 have to be moved. However, it should be noted here that the scanning device could be subjected to excessive loads due to the ongoing vibrations or that a cable break or damage to the optical fiber could occur due to the frequent and rapid movements.

A further alternative is to dispense with the lower scanning half 34 and instead to let the positioning device 31 rotate the pen through 180 degrees. This version is all the more advantageous since the lower half of the scanner tends to become more contaminated than the upper half due to the falling mine dust or paint particles.

6

The function of the positioning unit is explained below and preferred forms are described.

The positioning unit 3 shown in FIG. 1 serves to correctly position the pencils for image acquisition.

This is to relieve the data processing unit 7 by largely relying on the preprocessing steps such. B. the contour determination, pencil tracking and standardization measures can be dispensed with. The application of these algorithms is superfluous, since with the positioning device each pin is brought into the same and likewise known position - apart from the rotational position - at a known point in time.

It remains to be examined whether the pins are only attached to the same position in the scanning unit whether the same rotation position for everyone Provides a positive influence on the quality of the samples and thus on the classification result. So the position of the positioning device could be to turn each pin so that z. B. the embossing or another, clearly identifiable for each pen Always point up. This ensures that each Scanning cell practically always the same place on a pen measures. The positioning procedure can be used as a control loop describe. In the learning phase, a clear and with Little effort automatically recognizable positioning mark certainly. As an explanation, it is assumed that the Pen then assumed its prescribed rotational position if this marker is on top. The scanning unit acts as a measuring sensor of the control loop. With help The measuring cells become a rough classification algorithm queried after successfully positioning this marker should capture. The pen is now turned so long until the positioning mark has been recognized. In connection this can be fine-tuned. The Rotation to a certain starting position is at least for hexagonal pens due to their six stable layers less problematic. However, this requires positioning additional time and, if necessary, due to time constraints control algorithm supported by hardware.

The advantage of this positioning comes in particular  to bear when the characteristic values of the individual scanning cells are subject to large fluctuations, so that the correction each measuring cell for comparative assessment of the measured values different scanning cells is connected with great effort or a high percentage of errors remains despite the correction. The Correction factors would be too per scanning cell in the learning phase determine. Another advantage would be that for classification anyway necessary determination of the rotational position.

The positioning device could be used for fine adjustment to a grid with a rough scanning grid to be able to do without transformation as a corrective measure. As an example, consider a punctiform marking. Is the extent not larger than a measuring surface, the Marking in a scanning process as a point and thus be recognized correctly. The scanning grid will be half Extent of a measuring surface shifted in the x and y directions, so this point is made up of four measuring cells and each low intensity detected. This at different Position of the scanning grid (the so-called effect Sampling Theorem Explains) makes the investigation of imprints difficult and all other non-homogeneous places and occurs especially when an inhomogeneous area, such as. B. has an embossing or a printed image, has structures, which are no longer resolved by the scanning unit can. In this case, the samples can be of two correct embossing with different position of the scanning grid be differentiated from each other, that also the values of a blurred imprint are within this tolerance and therefore a classification is no longer possible. When fine positioning is It should therefore be noted that - for example with regard to embossing - Repeated fine positioning in the course of the scanning process would have to be made to all to be examined reliably classify inhomogeneous regions to be able to. If necessary, it may be necessary Positioning device so that it holds the pins  can also move in the longitudinal direction. The displacement route must then be at least the length of a measuring surface.

If the scanning grid is fine enough and a certain rotational position can be dispensed with, the positioning device is only responsible for positioning each pin that is to be scanned at a defined position in the scanning unit and storing it without vibration. In the case of hexagonal pins with a scanning matrix or scanning unit on both sides according to FIGS . 8 or 9, the positioning device could thus be constructed rigidly according to FIG. 10. In principle, only the lower receiving element 35 is required, since this forces the pin into a position that corresponds to one of its six stable positions anyway. The upper receiving element 37 serves to lock the pin, so that vibrations of the pin are avoided and the time is reduced which is necessary for a pin newly picked up by the positioning device to lie still.

An advantageous embodiment of the positioning and scanning unit consists in the use of two upper scanning halves opened downwards, which are arranged in parallel one behind the other. The scanning works according to the pipeline principle: the scanning halves check the half surface of a pen at the same time; a pen is only fully checked when it has passed through the scanning cascade. Each scanning half has a receiving bed, in which the pen z. B. is placed with the help of grippers. The receiving bed essentially takes on the outer shape of the pins to be tested. For hexagonal-shaped pins it has the shape of the receiving element 35 in FIG. 10, for round pins the receiving bed is correspondingly round. The test objects are pressed against the receiving bed for faster immobilization with the aid of nozzles and suction air introduced into the receiving bed. Compressed air is used to quickly release the pins after the scanning process has been carried out, and at the same time frees dust particles from the suction nozzles. The scanning unit is not movable. The scanning bed with the pressed test object can move upwards for scanning. The bed does not extend over the entire length of the pen, but is interrupted at one or two points. At these points, there is a gripper arm on the test object instead of a receiving bed. The center of rotation lies between the two holding devices. If the upper side has been scanned, the rigidly connected gripper arms (hereinafter also referred to as gripper arms for the sake of simplicity) transmit the test object to the other receiving bed. The test object is rotated about its longitudinal axis by 180 degrees, the underside can now be scanned as the top. Each gripper arm consists of two gripper fingers, their recesses being adapted to the shape of the pin. When grasping the pen to be checked, the upper finger grips, while the lower finger preferably remains rigid; when released in the second receiving bed, the gripping finger, which is now on top, lets go, while the upper finger, which is now on the bottom, preferably remains rigid and acts as a temporary receiving bed until the actual receiving bed is positioned accordingly for receiving the pen. Analogous to the receiving bed, vacuum can also be used with the gripper arm - if necessary only as a support. The gripper arm expediently moves downwards because the scanning elements are located at the top. For example, its trajectory describes a semicircle.

There are three constructive measures conceivable for the test objects to remove from the first bed and into the Take up second bed: 1: The admission beds move only in the vertical direction, the gripper arm guides a kind of linear movement for lifting out and putting in by. 2: The admission beds move exclusively in vertical direction, but can be removed and Dodge downward so that the gripper arm one describes a semicircular trajectory. 3: The first up  takes the fetch of the test object by a Backward movement and the second bed takes over The pen is deposited by moving forward after it is done Scanning process. When you pick up a pen, the pen can with the top just checked with a semicircular Movement to the second half of the scan while the second bed the pen just checked discards.

The proposed combination of scanner and posi tioning unit has the advantage that per transport cycle a pen is checked by parallel processing that the Positioning unit the rotation on constructive simple Way solves and that the scanning units through their opening be minimally polluted downwards. To reject Dirt particles can use compressed air that comes from in nozzles attached to the scanning unit slowly flows out. The The computing unit is responsible for determining the rotational position. Those adapted to the shape of the pen (e.g. round or hexagonal) Troughs in the receiving beds and the gripper fingers are manufactured as inserts and can therefore be quickly to be replaced.

7

The calculation unit is dealt with in this section. It possible classification features are first shown, then the data structure becomes a knowledge representation explained. This is followed by the acquisition of knowledge described. The description of the computing unit includes the description of the hardware.

As described in Section 1, a sample of Pins are analyzed in the learning phase and all are made of them necessary features extracted. The good-bad classification is only made possible by an upstream analysis. Out For structural reasons, it is obvious that the surface of the pen is in  Regions that correspond to the simple components subject to. Such regions are:

Homogeneous regions
Transition regions
Texture regions
Marking regions

The division chosen is not necessarily the only one possible or correct. However, it makes the most sense. When dividing the regions or choosing the As far as possible, knowledge should be known about the characteristics Structure of the pins are used. The more information flows in, the easier the classification is. However, the information taken as a help may not be the restrict general traps.

So it can be assumed that the pins are painted or be dip-coated, which means that in general sharp color transitions occur and the borders are more homogeneous Regions straight and also perpendicular to each other. It must be assumed that the positions of Embossments are very tolerant. Printed pens or printed areas of pens have one or more homogeneous regions of different color for the background, as well as a maximum of four basic colors for printing (e.g. Offset printing). The color of the natural wood is subject very strong fluctuations.

Homogeneous regions are regions of the same color that have no structure. Ideally, every job is with smallest deviation of the color intensity at the measuring wavelengths interpretable as a defect. The "good-bad" - Classification is based on the statistical parameters of the intensity of the measuring wavelengths to be measured.

The specification of an average as well as an upper and lower Tolerance limit (generally a value interval) that of  no measured value may be exceeded or undershot, provides the easiest way, which is also on a based on the theoretical model, which is the determination of a Probability of error with known distribution density allows. Scatter the measured values according to a Normal distribution, so the mean gives the expected value and the tolerance limits depend on the variance. Becomes e.g. B. as the upper or lower tolerance value three times Assuming variance, misclassification occurs of the homogeneous surface in question with a probability of 0.3%. Expected value and variance are from the To estimate the measured values of the learning sample. The parameters of the Distribution should during the classification phase e.g. B., constantly adjusted by moving averages and be improved, which increases the likelihood of a Misclassification reduced.

A homogeneous region is classified by Query each measuring cell that has a position within it homogeneous region measures to be contained in the specified Interval of values. Since the characteristic values of electronic components, the z. B. CCD elements or photodiodes are usually subject to very strong scatter, can the measurement errors z. B. reduce by determining a separate value interval for each measuring cell or each measuring cell is calibrated in the learning phase.

The transition region includes the transition from one homogeneous region into another homogeneous region. Possibly is also an extension to transitions between homogeneous and non-homogeneous or only between non-homogeneous Regions make sense.

The introduction of transition regions is proving to be helpful for handling the manufacturing tolerance cause different region dimensions. An additional Source of error, in addition to the discretization to be performed  to fluctuations in the dimensions of the sampled Leading regions is given by positioning, especially the raw pins already have tolerances. That is, a hexagonal-shaped pin is not arbitrarily precise in one its stable positions can be positioned. Since the scanning grid has only fineness, the measuring surfaces of the measuring cells cannot be made arbitrarily small, the Case occur that measuring cells whose measuring surfaces are in Transition between two homogeneous regions, one Deliver value that is not one of the expected values for the two corresponds to homogeneous regions. Rather, this mixed value depends depends on what the percentage of the two is Regions on the measuring surface. So it always becomes measuring cells give the values that indicate a defect let, although the transition from two homogeneous regions and the regions near this transition are also flawless unless this phenomenon is dealt with separately.

There are six types of transition regions that stand out distinguish the direction and number of their transitions:

Left-right transition
Top-bottom transition
Left-right-up transition
Left-right-up-down transition

If the scanning grid is too coarse, in which the width of homogeneous regions in the order of the grid spacing it may still be more advantageous, including those narrow regions as their own transition types in the To include transition regions. Set transition regions are made up of marginal areas of homogeneous regions as well as from small or narrow homogeneous regions.

A transition region must be chosen so large that the Transition of the regions concerned entirely within this transition region. Otherwise, an adjacent one  homogeneous region classified as faulty region. The location of the transition is within the transition region Tolerance-dependent and therefore not known and is normally allowed not be determined beforehand. However, can Information about the width and the course of the transition (e.g. sharp, flowing). A sharp transition should take place within a measuring area. At the Transition from the painted pen body to the tip of the pointed one On the other hand, pens run over several Measuring surfaces. In addition, this transition is on flat surfaces, as they are with hexagonal pens, parabolic pronounced.

Furthermore, additional information can be assumed, so e.g. B. that the transition from a (painted) region to a other (painted) region either horizontally or vertically runs. This means both with sharp and with smooth transitions that the measured values of the scanning columns or scan lines in which this transition takes place, have to be similar. Your spread interval can be as another characteristic of the transition in the learning phase to calculate. Is a tendency to shift the measured values recognizable from one direction to the other, this is how it runs Transition at an angle and there is committee.

There is a continuous transition across several Measuring surfaces, the values of the measuring surface in which this transition takes place towards the transition one gradual progression from one color to the other color of the have relevant homogeneous regions.

The transition regions are still larger than the transitions itself, so that - unless the transition is straight lies on the edge of a transition region - the left and right or above and below the transition regions Have homogeneity checked. The homogeneity criteria correspond to the neighboring homogeneous regions concerned.  

Texture regions are structured regions that generally contain multiple colors. Texture regions become Characteristics extracted without the importance of these regions to know yourself. So you will not z. B. in simpler Disassembled components and interpreted them, whereby a Overall description of the texture emerges. Understanding of Textures are left to the human viewer.

The location of texture regions can be very tolerant, so it is advisable to indicate a region in which the textured region certainly lies. Otherwise one Classified fault. So it’s not necessary To define transition regions that transition from contain a homogeneous to an inhomogeneous region, because these transitions take place within the texture region. It is also not necessary to transition from an inhomogeneous in another inhomogeneous region. The two inhomogeneous regions are considered as a texture region.

It also makes sense to include the actual size of the To determine the tetural region. In contrast to the location is subject to the extent of these regions is normally only slight Fluctuations once from the influence of one too coarse Scanning grid on the extent of the recorded image the texture.

The samples then contain the complete information and can thus be reliably evaluated if the resolution the scanning matrix is so fine that significant subtleties the texture in the image can be reproduced (sampling theorem). In addition to high costs for the scanning unit, this brings also increased circuit and computational complexity, so that in practice with a loss of information when scanning must be expected.  

An additional difficulty is that the investigating textures usually inhomogeneous textures are, d. H. the textures do not consist of a structural element, that these textures by appropriate stringing together generated.

Nevertheless, it is necessary to detect texture errors divided into smaller regions and for each of these elementary regions Determine parameters to match these with those in the Learning phase to compare the parameters obtained. In contrast on the procedure for the Analysis of complex patterns is the division of the Texture regions in elementary regions predefined and can be independent of the structural properties and the meaning of the texture, so possibly actually existing simpler components of the texture.

For this decomposition there are fluctuations in the dimensions as well as the location of the texture region. The Elementary regions should be chosen so small that the structural differences within such an elementary region are so small that by averaging the measured values As little information as possible is lost about this region. The elementary region could be dependent on the resolution for example a size of 5 × 5 or 10 × 10 scanning cells have. If necessary, it also makes sense the elementary regions from texture rows or columns put together. A smooth transition to the neighboring one homogeneous region can be treated by in a Preprocessing phase of texture classification scan lines or columns or parts thereof (especially beginning or End of the texture lines or columns) synchronization with that serving as a basis for comparison Get texture model. The may prove to be Use of row or column based feature extraction as beneficial.  

One method for obtaining features is determination a value histogram with the corresponding tolerance limits for each elementary region. The spread of values the histograms can be essential if the scanning is insufficient depend on the position of the scanning grid. A positioning unit, which is a fine positioning both in rotational as well as in the longitudinal direction (see section 6) is helpful in that the tolerance values are simple and reliable - albeit time consuming - in the learning phase could be determined. If necessary, the No need for fine positioning in the classification phase.

A defect is present if one or a few Elementary regions show significant deviations while the characteristics of all other elementary regions within the as a well accepted range of values.

Additional information (about the rough structure) of the texture If the scanning grid is too coarse, you can gain if there is a connection between changes in the parameters of the elementary regions (e.g. the histograms) depending on the Location of the scanning grid is produced. So every shift of the scanning grid in a certain direction tending to shift the values of the histograms or other characteristics of each elementary region. It applies to record this tendency as a function of the position of the scanning grid. For a texture to be examined, an elementary region and its analysis values (i.e. their Features) determined the position of the scanning grid. These Location is based on the mentioned function with the rest Elementary regions checked. There are significant ones Differences between theoretical expectations and the practical measurements, it is expected that the texture has a defect.

There is a significant deviation, for example, if one or a predetermined number of  Parameters by more than a certain amount from the expected The setpoint of the parameter or parameters deviates. The maximum permissible deviation is e.g. B. experimental determined and is generally dependent on the manufacturing tolerances. For example, B. a maximum deviation 20% of a parameter from the setpoint. For example, it could also be stipulated that a Inadmissible deviation is given if the deviation of a parameter or the weighted average of the deviations all parameters of an elementary region from the setpoint more than 300% of the average deviation, for example the parameter of the elementary regions recognized as well of their target values.

Other texture features are e.g. B. how quickly the Structure changed within a region. Such a thing The measure is approximately the sum of the changes in the sample values of one measuring point to another, where necessary making a distinction in different wavelengths as well as different Directions. An easy-to-calculate texture feature forms a histogram for each elementary region. For each measuring wavelength, the number of in measured values falling within a certain value interval.

In summary, the analysis of texture regions sufficient free scanning grid following advantageous hierarchical Approach suggested. The potential Region in which the texture to be classified is certain lies, is known and is determined in the learning phase. In A first step is to determine the actual Lay a texture of scan lines and columns of the potential Region characteristics determined. This can be, for example Histogram that, for example, only refers to the extreme values limited with respect to each measuring wavelength. in the need first step - from a corner point of potential Region outgoing - only the columns and rows are processed until both the horizontal and the vertical  Edge of the texture was found. The second step is processed every elementary region; characteristic values are determined and stored in a memory. The procedure for Determination of the characteristic values is fixed (e.g. histogram determination) or was in the learning phase due to certain characteristics the texture (rough or fine, symmetrical pattern etc.). In the third step, which only applies to textures is required, which extends over the entire circumference of the pen extend or that have no distinctive edges determined the rotational position and carried out a position normalization. After the position normalization are in the fourth step the parameters of the elementary regions using the above compared the mentioned comparison criteria with the target values. If there is a significant deviation, it is a defect classified and the process is ended. in the fifth step are the parameters of neighboring elementary regions summarized according to a specified regulation and compared with target values according to a further rule, with the tolerance limits being narrowed. The parameters obtained in the fifth step are analogous to this further compressed in the sixth step. The procedure breaks at a specified number of compaction steps from. The position standardization was the third 68110 00070 552 001000280000000200012000285916799900040 0002004132754 00004 67991 processing step assigned. You can without violating the proposed Principles also in the first, fifth, sixth or each further processing step.

A circuit arrangement for carrying out texture classification particularly quickly is explained below. The circuit arrangement consists of a control unit for sequence control, a sample memory for recording the measured values, a result memory for recording the results of the elementary region classification, a reference memory for recording the target values and a large number of hierarchically arranged computing units. The principle explained in FIG. 21 is used for the row and column treatment for determining the position of the texture region and for classifying the homogeneous edge areas. In the first processing stage, the characteristics of each potential elementary region are determined with the aid of a separate computing unit, which is activated by the central control unit and supplied with the required parameters (e.g. limit values, parameters or type of computing rule). The features obtained from the measured values in the sample memory and the comparison results of the features with the reference values stored in the reference memory are stored in the result memory. Analogous to the first processing stage, there are computing units in the second processing stage above, which - controlled by the control unit - access the results of the computing unit of the first processing stage in the result memory and carry out a further compression according to a possibly different computing method and the resulting characteristic values and those by accessing save the comparison results provided in the reference memory in the results memory. The structure and mode of operation of the third and each further processing stage corresponds to the structure and mode of operation of the second processing stage.

The control unit can access the result memory, to query comparison values and additional characteristic values to determine. The control unit only activates the processing stages required for the region to be analyzed as well as the required processing units. Lies For example, a texture region that is 100 × 100 Measuring points are scanned and have the elementary regions an extension of 10 × 10 measuring points each and are in the second processing stage the parameters of 10 × 10 each Elementary regions further compacted, so only the 100 computing units of the first processing level activated, which refer to the relevant, the samples of the texture region corresponding, refer to memory cells.

The control unit is also responsible for determining the position and the position normalization. The rotational position can also already addressed by the positioning mark - if available  be determined. The position standardization corresponds essentially a linear coordinate transformation what the determination of Offset addresses of memory cells means. For example the measured values or the characteristics in a memory matrix filed so that scan lines or lines of elementary regions are mapped onto lines of the memory matrix, so the normalization of the rotational position corresponds to the determination an offset value from the line address that is the start line or the starting line of elementary regions of the texture. The offset value can e.g. B. the row address itself or be the serial number of the line, describes the Moving the actual start line from the first Row of memory and is accessed when accessing the memory matrix used. The transformation can affect either the sample memory or the part of the result memory refer to the results of the first processing stage records. In the first case there is a more precise transformation possible by redistribution into elementary regions. For this purpose, the calculations of the first processing stage again.

The embossing region is a special case of the texture region. The investigation of embossing regions is facilitated due to additional information that can be assumed, or the quality of the results achieved with it is improved. An embossing region usually consists of characters and a color is generally used. The stylus, the development of which is shown in FIG. 11, has an embossing imprint that provides two embossing regions. If all the characters used in an embossing are of a uniform font, font-specific features can be used. Embossing regions are made up of simpler components. With a fine resolution of the scanning matrix, the areas of the simple components can be recognized and separated by z. B. areas perpendicular to the font are searched that do not contain a measuring point that contains the embossing color. It can be assumed that there is a relationship between several embossings that lie on the same side of a pen, since these are produced in one operation and with an embossing stamp. If two of the embossments are slightly to the right, the other will also be shifted to the right. In addition, the distance between such embossments will have little tolerance.

A faulty, i.e. H. e.g. B. blurred, embossed for example, if the character found does not match the expected character or a separator of two characters does not match the expected one Place.

If the resolution of the scanning grid is not fine enough to match the Areas between each character reliable and To be able to determine with certainty, so must the general Methods for the analysis of texture regions (see above) were used will.

The following procedure proves itself in the embossing analysis as particularly advantageous. It was in the learning phase how discussed above, the minting region in elementary regions divided. As with the texture regions, also in the learning phase determines a region in which the (error-free) Embossing is certain.

The elementary regions usually extend over the entire width of the texture region, but need not necessarily Form rectangles. For example, the letter A is a rectangular or trapezoidal elementary region be assigned.

Special properties of the elementary regions are in the form of characteristics expressed and in a data structure (see below) saved. Such a feature can e.g. B. a color histogram be. For example, B. each measuring wavelength specified  what percentage of the samples of the elementary region a relative intensity value of at least z. B. 80% have. Another feature that is particularly suitable for this specifies how many neighboring measuring points of the elementary region have the same intensity value. This characteristic represents is a measurement number for the characterization of lines, filled Areas and closed lines. In the case of Letter H, whose elementary region z. B. by the smallest If there is an enclosing rectangle, this will exist Characteristic from three details: a measurement number for the area requirement of the line as well as two details for the size of the Background above and below the bar.

The elementary regions are differentiated between filled and empty elementary regions. Filled in elementary regions contain embossing color while empty regions only the background color is present. So can for example the italic lettering HB made up of two filled in Elementary regions exist that are inclined by the empty elementary region are separated.

The processing unit for embossing analysis consists of a control unit for controlling the process, a sample memory to record the measured values of the region in which the embossing region must be located, a reference memory for recording the elementary region description, a computing unit for elementary region classification and one Computing unit for determining the location of the embossing region or Minting region within the designated region. The Position determination unit operates row by row or column and determines the enveloping rectangle of the embossing region. The structure and the principle was already in the orientation discussed by texture regions. The regional analysis works column-oriented and processes the columns of the Mint region in order. The scanning columns are there perpendicular to the longitudinal axis of a pin. A complete The scanning column thus extends over the entire circumference a pen. When processing the columns, the in Elementary region breakdown specified for reference information  are taken into account and those in the reference information determined characteristic values for this elementary region or checked. If all columns of an elementary region have been worked through, can the control unit for this elementary region a partial classification result can be returned. If there is a fault, the control unit stops and reports a faulty control unit above it Mint region back. If regional borders run diagonally, see above can process several elementary regions at the same time be.

8th

Below is an example of a data structure for management the reference data obtained in the learning phase.

In the data structure, all information that is characteristic Appearance affects all pens of one sort is common, managed and stored so that for a pen to be examined as efficiently as possible "good" or "bad" statement can be made.

The data structure is in the learning phase for each pen type rebuild. That remains during the classification phase essentially unchanged. However, in the classification phase ongoing improvements to classification parameters can be made. However, this remains the shape of the data structure unchanged.

The structure of the geometric part of the data structure reflects the color structure of the pen surface. Fig. 11 shows the development of a hexagonal pencil with an embossing in white. Also shown is a scale that shows the proportions of the pen. The edges of the pen are painted black. The basic color is yellow. The two scales at the top and left edge of the development indicate the position of the scanning grid. The location of each scan cell, or the location of the stylus measured by that cell, is clearly determined by a pair of two numbers. The measuring center is the intersection of the relevant scale lines. Each intersection of scale lines corresponds to a measuring center; each measuring center represents a measuring cell. The size and the shape of the measuring surface is determined by the design of the scanning cells and their arrangement in the scanning matrix. For example, only the point of the measuring center can actually be measured or a very large round area that extends over several measuring centers. To simplify matters, it should be assumed here that the measuring surfaces occupy the entire pen surface covered by the scanning matrix and that the measuring surfaces are also disjoint. This requirement is only met to a limited extent by the usual measuring cells. Rather, the measuring center is normally weighted highest, while the edge of the measuring surface, which is still significant in the measurement result, cannot be determined exactly due to the lower weighting.

FIG. 12 shows a possible region division, as it could have been determined in the learning phase. Analogously to FIG. 11, the scanning grid is also shown here. The numbers assigned to the individual regions should clearly identify them. The region shown in broken lines and labeled "NP" represents the zero position (positioning mark). It is necessary to determine the angular position of the measured pen. As a hexagonal pin is used as an example, the zero position is not determined by specifying an angle of rotation, but only by one of six stable positions.

In the Fig. 13 and 13a a possible data structure for the representation of knowledge determined in the learning phase is provided. The data structure of Fig. 13 consists of four blocks of information. The data are accepted and are not based on the evaluation of a specific pen type. In the "General Information" block, general data is managed, such as B. the number of colors recognized in the learning phase, the number of physical channels that are used to distinguish these colors, the number of patterns of the learning sample, the number of patterns that contributed to parameter improvement during the classification phase, etc. For reasons of efficiency, the The number of physical channels is usually equal to the number of measuring wavelengths (see section 4). There are no strict rules about what general information should be stored. However, the principle applies that the more information is stored, the easier and faster the required information can be calculated or made available. For example, a type of type designation or a list of characteristic data can be specified for the pen type to be examined. B. provides information about the length, shape (hexagonal, round, oval) or the quality of the pen. Pens that are lacquered at the end are usually more expensive than pens that do not have this lacquer finish. Different quality requirements may be placed on different types of pens.

Another block of information contains all the necessary data about the occurring and distinguishable colors. In Fig. 13, four colors (measurement wavelengths) were considered sufficient for color discrimination. The colors are identified by name, although the color designation has been included for the sake of clarity. In fact, the colors are not known to the system by name, but can also be identified by numbers. The color names are only of interest if information about the examined pencils is to be communicated to quality assurance. So from the system z. B. statistical information on the frequency of errors and the location of the errors in the form of measurement reports can be collected and printed. In this case, colors can be saved in a system table with the associated characteristic spectral characteristics. A measured spectral characteristic is assigned to the color whose spectral characteristic comes closest to this.

In FIG. 13, the spectral range is measured at the four positions 460, 500, 588 and 710 nm to differentiate the colors. The mean value and the upper and lower tolerance limits are given. Possibly there is no misalignment, so that the upper and lower tolerance values are the same. More precise spectral examinations will show whether even a normal distribution can be assumed. This would give you the opportunity to control the error rate of the system in a controlled manner.

The parameters of the zero position are specified in the third block. So it is specified in which rectangle this is with certainty. Also, one or more Characteristics of this zero position are saved.

In the geometry block, which is shown in Fig. 13a, each region is managed in a respective node of a graph which shows the neighborhood and contain relationships of the regions. The classification parameters are saved for each region. In FIG. 13a, this is, for example, the rectangle in which the region is located, the type of region, etc. In the case of transition regions, the color of the relevant homogeneous regions (left = left, right = right, o = top, u = bottom, lo = top left, ro = top right) and a number indicating the number of scanning cells over which the region transition extends (1 or 3 for the transition to the untreated wood at the top). For texture regions, the location of the texture regions and their size are saved. The frequency of occurrence and a tolerance interval are stored for each color. The scheme could be simplified by managing one record for each different region. In Fig. 13a there are only two different homogeneous regions, but 12 data sets for homogeneous regions are stored. Simple features have been used in Figure 13a. The features mentioned in the text or combinations thereof can be used equally.

9

The knowledge acquisition is explained below. Introductory there is a brief description of the learning phase. It will learning in homogeneous regions with regions of origin considered. This is followed by a description of learning for texture regions. The conclusion of the description the knowledge acquisition form explanations for determining the Measuring wavelengths and for adapting knowledge.

The knowledge acquisition of the system extends to both the learning phase as well as the classification phase. In the Learning phase must be based on a sample of pens characteristics to be determined for classification, during the course of Classification phase a constant improvement of the characteristics and adaptation to changed environmental conditions becomes. The first way of acquiring knowledge here is as learning and the other called knowledge adjustment. Frequently a learning sample with flawless pens used to to make the learning phase as easy as possible.

The learning phase can be structured according to FIG. 14.

First, a pencil pen of the learning sample is scanned, the spectral characteristic of each measuring point being recorded. To save time, the spectral characteristic is measured, for example, at 20 wavelengths or wavelength ranges.

An improvement in the quality of the features can be achieved  when the pin is in different positions by means of the positioning unit Layers are brought and measured again. The influence of the scattering of the characteristic values of the scanning cells reduced, d. H. it will ultimately be a calibration of the scanning cells.

After preprocessing measures such. B. Frequency evaluations or threshold value operations, the classification features discussed in section 7 are determined and stored in a data structure (see FIGS. 13, 13a).

All remaining pens of the learning sample will be the same analyzed, being during the analysis of the first Pen received information can be accessed. These further analyzes serve to improve the characteristics and thus increasing the reliability of the classification.

An essential one, but not necessarily necessary, but nonetheless, quality-improving support would be given, if the system before the actual learning sample too Pass samples of single color pens for calibration purposes the colors are only the colors of the examining pen type. It would be proportionate simply, the measuring cells with the help of the Fully calibrate the positioning device for each color and determine those wavelengths that are sufficient around the individual colors reliably and quickly to be distinguished from each other. In this case, could - based on the already known colors a more reliable Region classification can be made.

The computing effort in the learning phase can be reduced for some Shorten pen types if all previously discussed pen types stored in a database. So with that first pen in the sample determined whether the pen type is already known. If this is the case, they can already  stored characteristics as a basis for comparison will. Give way to the features of the new pen series not from those of the old series, so the classification phase with the more reliable old classification parameters be started. Otherwise, by averaging to generate new features. If the deviation is too great however, the old characteristics must not be used will.

If there are only homogeneous regions, then - after determination the discretized spectral characteristic of each measuring point - the location and extent of these regions Threshold operations are determined. For this, the discretized spectral characteristic of each measuring point with the Characteristic curves of its neighboring points compared. Exceed the deviations a fixed limit, so there adopted a regional boundary and it is in the data structure entered a transition region. Otherwise you are within a homogeneous region. The Know that the region borders only horizontally or vertically run, can be used to check the plausibility of the received Results are used.

To compare spectral characteristics, a distance measure is required that indicates how different two spectral characteristics are. The maximum distance (L _∞ standard) indicates the maximum difference between two functions, while the Manhattan distance (L₁ standard) relates to their mean deviation. The maximum standard thus weights peak deviations and suppresses smaller deviations, while the Manhattan distance - at least for many measuring wavelengths - completely suppresses peak differences with a small width. On the other hand, the Euclidean norm takes into account both the average deviation and also emphasizes maximum differences, since these are squared in the norm. The Euclidean distance requires multiplication; the root calculation can be dispensed with. Squares can overflow 16-bit numbers. In order to be able to work with numbers of this length, a standard is available as an alternative, which forms the mean of the maximum standard and the Manhattan distance.

The threshold of the norm, which determines when two spectral characteristics are the same or different, heuristic fixed. In addition, it is possible to use this Adjust the value in the course of the analysis. Is also a calculation the threshold value with a frequency evaluation feasible on the basis of the distance dimension. To this The distances from as many spectral characteristic pairs as possible are used calculated and the frequency of occurrence of the individual difference values are counted. With this one-dimensional Frequency analysis should be a narrow and very limited dense range of values occur between 0 and a number x to be determined. x can then be used as a threshold be used.

Once the regions have been divided, each can Color by averaging over the characteristics of the measuring points a characteristic of the homogeneous region in question Spectral characteristic can be determined. Each of these spectral characteristics specifies a basic color. It should also the spectral characteristic of the color of the untreated wood to be available. These characteristics then become those Wavelengths determined that are sufficient to the individual Differentiate colors from each other safely.

The evaluations described must at least for the first pin of the learning sample. For the rest Pens, these evaluations can be repeated. It can but also on some of the analyzes, such as B. the determination the threshold value or the measuring wavelengths and you still get stabilization the values of the characteristics. The results should always be with  Help with plausibility checks (e.g. sharp and straight Transitions between regions) checked for correctness will. However, should the analysis of the second Point out a significant deviation, so are Repeat all evaluations and the results with the Average values of the first pen. This is how it works long repeated until the learning sample has been processed. One way of checking is to look at each pen classify the previously determined information.

Marking regions normally only take a small part (e.g. less than 50%) of the homogeneous region in which they are lie, so that the embossing due to excessive fluctuations of the measured values as a defect and thus as an inhomogeneous region can be recognized and delimited. However, this assumes that the pens of the learning sample are error-free.

One could check the functionality of the Device also after the learning phase use a negative sample. The system should then show the user which pen is recognized as reject as well as the cause of the error and the location of the mistake.

The following paragraphs relate to learning at Texture regions. The problem is first explained. Then four solution methods are discussed.

Data reduction using cluster analysis,
Geometric solution method,
Solution of a system of equations as well
Search for a linear independent basic matrix.

The analysis of texture regions is then problematic, if the spectral characteristics of the primary colors are not  are known. These form the basic information for the System learning algorithms.

Are there only homogeneous regions that are due to the Fineness of the scanning matrix can be resolved the basic colors are also known according to the regional division. Color mixtures, d. H. Pseudo colors that arise when the resolution of the scanning grid is too coarse, only come in the transition regions. These are therefore used for determination discarded the spectral characteristics of the primary colors. The Color of an embossing can be recognized using simpler methods, because embossments consist of one color and the basic color the homogeneous region in question is known.

Texture regions cannot be easily determined whether the spectral characteristic of a measuring point is a pseudo color mixture is made up of two or more primary colors, or whether this characteristic curve represents a basic color.

However, it is necessary to add the basic colors of a pen determine. These have the "extreme" spectral characteristics, from which the measuring wavelengths are calculated. For clarification imagine 100 measuring points, with two Basic colors are present and only one cell has the spectral characteristic of the pure basic color. All other 98 characteristics have due to the non-punctiform expansion of the measuring surfaces different percentages of the two Colours. It can therefore have 100 different spectral characteristics give. If these are sorted accordingly, one can smooth transition from the spectral characteristic of one Determine color in the spectral characteristic of the other color. It is clear that the calculation of the measuring wavelengths so, those wavelengths that make the distinction of the two colors allow to be reliable only can, if this on the spectral characteristics of the two Primary colors. The to these "extreme" characteristics belonging basic colors are also called corner colors. About that  in addition there are indeed primary colors that are a mixture of colors are from corner colors.

Finding the spectral characteristics is easier, ever the resolution of the scanning matrix is finer. Are the first Pens of the learning sample painted in one basic color each, this eliminates this problem.

The following explanations deal with the finding the spectral characteristics of the corner colors. At this Place the mathematical formulation of the problem introduced. It is not meant to be continuous, but rather be worked with the discretized spectral characteristics. Spectral vectors are therefore also spoken of. The following notation is agreed:

M number of measuring points
N number of corner colors
K Number of nodes of the discretized spectral characteristics
w ⁱ spectral vector of the i-th scanning cell
w _j ⁱ j th component of the spectral vector of the i th scanning cell
W ⁱ spectral vector of the i-th corner color
W _j ⁱ j th component of the spectral vector of the i th corner color

The problem is first explained using two primary colors. Given M measuring points and thus M discretized spectral characteristics w ⁱ (i: = 1,..., M), each consisting of K measured values. It is known that there are only two primary colors. Assuming that the characteristic curve of a mixed color is created by linear superposition of the two primary colors, any characteristic curve of any color mixture of these two primary colors can be represented as a linear combination of the spectral characteristic curves W 1 and W 2:

w ⁱ = t ⁱ W ¹ + (1-t ⁱ ) W ²
0 <= t ⁱ <= 1, i: = 1,. . ., M (1)

t ⁱ indicates what percentage of the measuring surface i is wearing the color 1. Accordingly, the measuring surface consists of (1-t ⁱ ) 100% of the color 2. The w ⁱ and the W ⁱ are vectors of length K. Thus, (1) describes a straight line in K-dimensional space with variable t ⁱ . The straight line is finite. Its key points are W ¹ and W ². All variables on the right side of the equation are unknown. These are M + 2K unknowns. Overall, (1) consists of MK equations. Only the vectors W 1 and W 2 are of interest. The problem can only be solved with two primary colors if the spectral vectors of the corner colors are present in the set of vectors w ⁱ (i: = 1,..., M). It is important to find out the two spectral vectors from this set.

If, however, there are N corner colors, this is written System of equations too

The right side of (2) describes with variables t _j ⁱ an N-1-dimensional hyperplane in K-dimensional space (provided that K <= N). Due to the condition Σt _j ⁱ = 1, this is limited.

Normally, many of the spectral characteristics mentioned above are of the sampling points are identical. This then lights up if you take into account the number of sampling points. So could a pen z. B. can be scanned at 5000 points. The results are multiple spectral characteristics of the algorithms described below, but rather due to additional rounding errors worsen. That is why - not least  Time constraints - a reduction in data volume required.

A method is given here that groups of the same or almost identical spectral characteristics to a medium one Summarized spectral characteristic. If necessary still noted the number of characteristic curves leading to such a Belong to cluster point. A distance measure that the deviation between two characteristics was determined in the discussion of learning in homogeneous regions.

The procedure for frequency analysis of spectral characteristics builds a K-dimensional adjacency graph. Every knot of the graph corresponds to an accumulation area. Every knot or each cluster area refers to those cluster areas, that are in the immediate vicinity. In addition to The distance becomes the thickness of the cluster area these neighboring areas in the node.

For a newly added discretized spectral characteristic the node of the adjacency graph is searched for whose Spectral vector comes closest to the new one. Falls short the difference between the two discretized characteristics given size, the vector is assigned to this node. This is done by averaging with appropriate weighting determined a new overall characteristic and the number of corrected vectors belonging to this node. Further the distances to the neighboring nodes must be adjusted. Is the distance of the spectral vector from that determined If the knot is too large, a new knot is entered.

A geometric solution method is described below.

As explained in the introduction to learning in texture regions, the right-hand side of equation (2) describes an N-1-dimensional hyperplane in K-dimensional space. This level is in turn limited by N-2 dimensional hyperplanes. The intersections of N-1 such planes, ie the corner points of the N-1-dimensional hyperplane, are the vectors sought. These are the spectral vectors of the primary colors. In other words, the corner points are characterized in that all but a t _j ⁱ in (2) become zero.

Fig. 15 shows the spectral characteristics of possible discretized points of measurement, the spectral characteristics are measured on 3 wavelengths. The characteristic curves are represented by points. There are no measuring points whose measuring surface contains exactly one corner color. In the following it should be shown that the spectral characteristic of the basic color can still be determined. The two-dimensional plane described by (2) with variable t _j ⁱ is also shown. Referring to FIG. 15, there are three corner colors. For the sake of illustration, the spectral vectors of the corner colors are orthogonal, ie their spectral characteristics are so sharp that each of these characteristics has a non-zero intensity component in exactly one measuring wavelength. As a rule, the spectral vectors of the primary colors are not orthogonal, but with a few exceptions they are linearly independent. The measuring wavelengths must of course be chosen so that they approximate each of the three spectral characteristics as well as possible and also distinguish the three primary colors with the highest probability of success. In Fig. 15 it is further assumed that the individual vectors are normalized.

The spectral characteristics should lie in one plane due to the conditions Σt _j ⁱ = 1 of (2). In fact, this will not always be the case, which is due to rounding and measurement errors.

The edges of the spanned by the spectral vectors Level consist of spectral vectors that are made up of the mixture  of the relevant two corner colors. Internally vectors are color mixtures of all three corner colors, d. H. Linear combinations of their spectral vectors.

It is not known how many of the basic colors there are, still look like their spectral characteristics. However, they are Corner vectors of the spanned plane are the spectral vectors of the Corner colors. Hence the number of these corner vectors is Number of corner colors.

The procedure for determining the corner points is such that the convex hull of the spectral vector set is first calculated. This step is shown in FIG. 16. If the set of spectral vectors is pre-processed with the cluster analysis (see above), the determination of the convex envelope can be supported by means of the graph discussed there.

All vectors that do not belong to the convex hull are discarded. A linear interpolation is carried out with the remaining vectors, only a certain approximation error being tolerated. For this approximation, two spectral vectors neighboring on the convex hull are first taken out and a straight line is put through these two points. If the approximation error to one of its neighboring points exceeds a defined limit, a straight line was found. On the other hand, the interpolation line is determined by these three points and the process is continued iteratively. The result of this smoothing operation on the convex hull is shown in Fig. 17.

The approximate convex hull of Fig. 17 now consists of 6 straight lines. In the next step, those lines are eliminated whose lengths are small compared to the larger lines. The remaining straight lines are cut together. These lines limit the calculated plane spanned by the spectral vectors and their intersections are assumed to be the spectral vectors of the corner colors.

In the example of FIG. 15 there are only three measuring wavelengths. However, so many measuring wavelengths should be used to reproduce the spectral characteristics of the sampling points with little information loss (sampling theorem). This increases the computing effort. In particular, if there are more than three corner colors, it is not straight lines, but planes or hyperplanes that should be approximated. The discretized spectral characteristics of the corner colors cannot be calculated exactly if there are no or only insufficient spectral vectors at the edge of the actual hyperplane.

In addition, the pen can also be painted in colors, are the mixed colors from the corner colors calculated above. This results in the spectral characteristic of such primary colors as the superposition of spectral characteristics of the corner colors. Therefore, these primary colors cannot be derived from the pseudo colors distinguish that arise because more areas of primary colors lie in a measuring surface.

A reference to such "inner" primary colors can be obtained with a frequency analysis (see above). If a closely spaced cluster of points is detected, it can be assumed that the vectors of this cluster represent a basic color. This presupposes that these spectral vectors do not belong to transition regions. Such spectral vectors should be excluded from the frequency analysis anyway. In addition, it will hardly be necessary to apply the method mentioned here to mine pencils that only consist of homogeneous regions. Rather, it is to be used for texture regions. According to the above considerations, the dot cluster from FIG. 15 is assigned to a basic color.

As a further alternative, the solution to a System of equations shown.

The determination equation (2) for the w ⁱ (i: = 1,..., M) represents a system of equations with M equations. Since the W ⁱ are vectors of length K, each equation consists of K individual equations. Thus there are a total of KM equations. The N vectors W ⁱ (i: = 1, _... , M) and t _j ⁱ are unknown. Because of the conditions Σ t _j ⁱ = 1 these are (N-1) M unknowns. The system of equations thus contains a total of NK + (N-1) M unknowns in KM equations. If the condition KM <NK + (N-1) K is fulfilled, the system of equations is overdetermined, unless linearly dependent equations occur.

Because all unknown variables on one side of the system of equations are not explicitly solvable. By skillful reshaping it can be brought into a shape so that the linear approximation problem is an iterative Solution allowed.

The system of equations is practically impossible to solve, though not the spectral vector set beforehand using cluster analysis was significantly reduced.

However, it is expected that the matrix of this system of equations is poorly conditioned, so not always one Solution exists. Consider the case for two basic colors. For K = 2 nodes and four spectral vectors this should be solvable. In fact, this system of equations always unsolvable.

In contrast, the proposed geometric method would find a solution here too. Because of the formation of the convex The spectral vectors that are the largest are selected Distance from each other. The geometric process has the advantage that its result by combining the theoretical approach with heuric considerations crucial can be improved and stabilized. It's closed expect the solution to the system of equations - if so is at all solvable - consists of the spectral vectors that form the edge of the convex hull. These are geometric First step procedure with probably less  Effort calculated.

Furthermore, as explained below, the spectral vectors can be of the corner colors using a linear independent Describe the basic matrix.

As already mentioned, all spectral vectors are as Linear combination of the spectral vectors of the primary colors (i.e. of the corner colors). The spectral characteristic of a However, corner color does not result from the superposition of the Spectral characteristics of the remaining primary colors.

This reduces the finding of spectral vectors from Corner colors on determining a linear independent Basic matrix. All vectors of this matrix are the spectral vectors the desired corner colors.

The procedure could be such that initially from the with the help of cluster analysis reduced amount of spectral vectors two vectors are selected, the largest Distance from each other. Now this crowd becomes selected another spectral vector and checked whether all three vectors are linearly independent. Is this the Case, this is included in the set of spectral vectors potential corner colors added. Linear dependency can be demonstrated with the determinant, which is then zero - or at least almost zero - is. Already be a lot of such linearly independent vectors. A new one added vector that is linear independent of all other vectors of this set is added to the set. However, if this vector is linearly dependent, then determines those vectors whose linear combination this Vector results. There is a possibility that this vector works better as a vector of one corner color than another, which is already a vector of a potential corner color was recognized. In this case there is an exchange step required. A vector is more appropriate if  this closer to a corner of the actually stretched and hitherto unknown level lies than any other. This then also means that this vector is outside the convex Envelope of the vectors of the selected set lies. In order to must have one or more coefficients of the linear combinations be negative. The new vector becomes the vector exchanged, the coefficient of which has the negative value. All vectors related to this new vector are linear are independent, have a coefficient that is almost Is zero.

This method is particularly suitable if of them it can be assumed that the spectral vectors of the corner colors in the set of spectral vectors of the measuring points are included.

Based on the spectral characteristics of a pen occurring basic colors must determine the wavelengths that are sufficient to ensure all spectral characteristics and to be able to reliably differentiate from one another. At the the simplest would be to complete the H. the discretized, Use spectral characteristics for comparison. Indeed this results in an unreasonable expenditure of time when measuring the characteristics of the individual measuring points as well in the classification. A restriction to certain Wavelengths are possible because some support points of the characteristics very little to differentiate the characteristics contribute from each other. Other support points in turn have a significant part in their distinction.

The selection of the measuring wavelengths is of crucial importance Importance. So the number of measuring wavelengths influences the Computing time. Furthermore, the position of the measuring wavelengths significant impact on the reliability of the classification and thus on the probability of failure of the system.  

So there is an optimization problem to solve, with a Quality criterion - the reliability - depending on optimized the number of measuring wavelengths and their location becomes. If necessary, the reliability for a fixed Number of measuring wavelengths optimized as a function of the position and - based on the quality criterion or a cost function - a tolerable number of measuring wavelengths selected. It remains to be examined to what extent it is necessary to specify a set of measuring wavelengths that distinguish each spectral characteristic from all others helps, or whether for any combination of spectral characteristics a separate set of measuring wavelengths is to be determined.

The selection process should basically consist of two stages consist. The first step is a lot of out of the question upcoming measuring wavelengths selected in the second step is reduced based on a quality criterion.

The easiest choice is the first Level when all extrema of the spectral vectors of the primary colors be determined. The potential measuring wavelengths are then all wavelengths at which at least one Spectral characteristic a local minimum or a local or global maximum.

A better result will probably be delivered if instead, the for all pairs of spectral characteristics Difference function determined and as potential measuring wavelengths selects those with at least one difference function has an extremum. The resulting one The amount of wavelengths should be larger than that mentioned above Procedure.

To reduce the number of wavelengths in the second Level is initially a measure of reliability or effectiveness specify. This could be a two-dimensional one for each wavelength Matrix, consisting of G × G elements, occupied  will. G is the number of primary colors. The ÿth field element records the difference, the i-th and j-th spectral characteristics or spectral vector at the point of this Has wavelength. If there are K wavelengths, then there are there K such matrices. The effectiveness of a wavelength is judged by how many of its matrix elements one exceed the upper limit and according to the number of Wavelengths used to distinguish a spectral characteristic from everyone else. It doesn't make sense that Select wavelengths that have a specific spectral characteristic differentiate very effectively from all others let other spectral characteristics with their help cannot be distinguished from each other at all. The purpose of the evaluation is to minimize the probability of errors. It can therefore also be useful in the quality criterion the probability of occurrence of the Include primary colors.

In addition to heuristic procedures, the problem can be solved formally describe and solve. In particular, a lead probabilistic approach to the goal.

Knowledge adjustment extends to continuous improvement the classification parameters during the classification phase. So the statistical parameters of the measuring wavelengths be continuously adjusted. It is suitable for this the moving average calculation. For stability reasons of course the new value is weighted less are considered the previous value.

The improvement could be classified by everyone Pen. However, the one with the knowledge adjustment associated time expenditure only slightly on the lead time of the system. It should be enough after to only match some parameters in some classifications.  

Improving the features is problematic, thus the measuring wavelengths. This is necessary if of them it must be assumed that the tone of the basic colors or for example the intensity of the embossing (e.g. the depth of impression) change over time. To this The purpose of the measuring wavelengths with the lighting unit be reset. Algorithms of the Learning phase - or at least complex calculations - called will. So it will be necessary to complete the classification phase to interrupt and switch on a learning phase.

10th

Due to the requirement for the shortest possible lead time the mine pencils - except for the learning phase - cannot be a conventional one Data processing system can be used. Much more parallel work is required, the corresponding Hardware based. The so-called pipeline mechanism delivers a means of increasing efficiency. At the Pipelining are the processing units that actually be called up sequentially, active at the same time. they edit data from multiple pens, changing the data are in different processing phases. So can e.g. B. a pen can be scanned while the samples of the previous pen are still in preprocessing and whose predecessor is just being classified. To this Three or more pens are being processed simultaneously. To implement the pipeline principle, the individual receives Processing levels some local storage.

The hardware structure is closely linked to the functional structure of the system. This Fig. 18 shows.

The system consists of four processing units, the Preprocessing unit, the unit for feature extraction, the classification unit and the unit for improvement  the parameter. The feature extraction is partial cannot be separated from the classification. So at homogeneous regions very simple characteristics, namely the Samples themselves. The effort becomes Feature extraction at the expense of hardware to a minimum reduced. A controller takes over the monitoring of the Overall process. It continues to have a decisive share in the learning phase. There algorithms are used that are not absolutely need the support of hardware, because in sufficient time is available for the learning phase. All processing steps make use of the methods. You can Algorithms or special circuits. Furthermore, must accessed the data structure stored in a memory will. In addition, the data structures of pens already examined in a knowledge base to get managed. This would reduce the effort in the learning phase to reduce.

If necessary, the processing sequence "preprocessing - Classification "if the Classification z. B. of textures only after a new one Scanning with fine adjustment is feasible.

Fig. 19 shows an example of the structure of the preprocessing unit. The signals of the scanning unit are fed into this unit. These are digitized with the aid of analog-digital converters 44 and buffered in buffer memory 45 . The sample values are stored in a buffer memory 46 , which can hold all sample values, ie vectors, of a pen. It doesn't matter whether the pen is scanned line by line or in a single step. A unit for position detection 49 determines the rotational position of the pin.

If necessary, the control unit for fine adjustment 48 can be used for fine adjustment of the measurement object. The samples are then copied into the pattern memory 47 after a position normalization. The buffer memory 46 can be identical to the pattern memory 47 , since the position standardization only changes the offset address, provided that the pattern memory can be addressed cyclically. If necessary, the samples are still calibrated. Fig. 19 is intended to show the calibration of the sensing cell at hex-shaped pins with the aid of a Kalibrierspeichers 50th Here, each measuring cell can only measure six different points on the pin because it only has six stable layers. The calibration factor of each measuring cell is thus stored for each position. Such a calibration can only take place after the position detection. The calibration memory is occupied in the learning phase.

Fig. 20 shows the structure of the classification unit prinzpiellen. The parallel work is carried out with the help of numerous, independently working processors, each of which solves a small subtask. The hardware of the entire system must be designed to be flexible in order to cope with the diverse tasks. The classification unit according to Fig. 20 divided areas of responsibility in the five control and analysis of homogeneous regions, the transition regions, textures and embossed. The control takes over the coordination of the four analysis units, which all work simultaneously and completely independently of one another and combines their results into an overall classification result. The areas to be examined, ie the regions, are determined in the learning phase and stored in the data structure.

Fig. 21 shows the structure of the processing unit for classification of homogeneous regions. The central unit is again a control unit 56 , which has its own processor. Furthermore, as mentioned above, this unit has its own pattern memory 51 . It is the task of the preprocessing unit to occupy this memory 51 . In this case, only those memory words need to be assigned values whose assigned sampling points belong to homogeneous regions. The processing unit shown in Fig. 21 assumes that the data in the local pattern memory 51 can be accessed line by line. The bars of the pattern memory 51 in FIG. 21 thus represent memory words of scan lines. The processing unit also contains a reference memory 55 . This is initialized in the learning phase on the basis of the data structure discussed. In principle, it can be structured like the sample memory, but contains more information per memory word. An interval of sample values is stored for each sample cell, as well as an indicator which indicates whether this cell belongs to any homogeneous region or not. For reasons of storage space, membership in a homogeneous region could be indicated by the fact that the permissible value interval is different from zero.

The classification is done line by line in parallel and sequentially within each line. For this purpose, a counter 54 addresses all words of each memory cell including the associated reference word. A comparison unit 53 checks whether the samples lie within the permissible value interval. If this is not the case for at least one value, the pen can already be classified as faulty. For quality control, the addresses of the scan cells that have incorrect values, including these values, are written to a log memory 52 . This could e.g. B. be a shift register memory. While a comparison unit must be provided for each memory line, the counter 54 and the error log memory 52 can only be present once. The values of the reference memory 55 can, if necessary, be adapted during the classification phase. The comparison unit 53 could make the adjustment by means of moving averaging. A value may only be adjusted if the relevant sample value of the pen is free of errors. The reference memory 55 presented in FIG. 21 can also be used if the same value interval applies to all samples of a homogeneous region, although in this case the memory structure could be simplified.

Similar to the classification of homogeneous regions the processing units for treating the remaining Regions subunits. If necessary, these are subunits multiple present. Each subunit could do the analysis of a region. Allocation of the Subunits for the individual regions are based on the data structure in the course - or at the end - of the learning phase.

11

The most important sub-problems are dealt with below and explained in the overall context.

A fully automatic system is proposed that examines the surface of all types of pencils and automatically sort out defective pencils. The system should be able to process several pens / second. To support quality control, statistical Evaluations are made and measurement reports are issued. The variety of pens is very large (over 1000). The system is able to adjust itself to new pens and recognize pen types that have been checked once. To this end, system activity is divided into two phases subdivided. In the classification phase, every pen examined and discarded as a committee if an error was detected. In the learning phase that precedes the classification phase, the system determines using a few faultless pens (e.g. 20 pens) all for a "good-bad" classification necessary parameters. The classification parameters are quite unreliable and because of this small sample  can only be roughly estimated. That’s why System's ability to do this during the classification phase to improve and thus stabilize. The additional Learning in the classification phase is also important because the environmental parameters change over time. So the electronic components age, temperature fluctuations change their characteristics and color tones can be subject to certain fluctuations.

According to FIG. 1, the system consists of a scanning unit 2 , which also takes over the homogeneous illumination of the pen surface. The scanning unit is made up of a large number of measuring cells which are arranged in a grid. The light required for illumination is supplied by an illumination unit 4 and fed into the scanning unit with the aid of an optical waveguide 5 . The scanning cells should be as close as possible to the pin to be scanned. There is also a positioner unit 3 , which holds each pin in a specific position in the scanning unit and, if necessary, can rotate the pins and shift them longitudinally by the size of the scanning grid. The transport of the pins is controlled by a transport unit 1 . Faulty pens are sorted out by means of an ejection unit 6 . The entire process is controlled by a data processing unit 7 . The system is said to either be coupled directly to the production machines or to be able to take over the pens from a separate pen memory. The pins must be able to be picked up in an orderly manner, ie the tip of each pin must point in the same direction.

The only physically measurable quantity is the intensity of the reflected radiation at different wavelengths or wavelength ranges at the measuring points. This information, d. H. the spectral characteristic for as many as possible Points of the pen must be determined to carry out the "Good-bad" classification of the pins is sufficient. Since the  complete determination of the spectral characteristics on all Measuring points and the subsequent evaluation are very time-consuming the system is limited to the selection of Measuring wavelengths. Usually represent the measuring wavelengths Spectral ranges that have different widths can. The spectral characteristic of all measuring points each Examining pen is therefore only for these measuring wavelengths measured. The measuring wavelengths are for each font and possibly also for different regions of the same pen different. The amount of measurement wavelengths is determined in the learning phase. There is one in there Optimization task. The position of the measuring wavelength must be the smallest possible number can be calculated so that a reliable differentiation from a defect possible is. The correct choice of these wavelengths is crucial Influence on the probability of errors in the system.

Since the measuring wavelengths differ from pen type to pen type, a multi-channel scanning device is required, and the wavelengths of their channels should be changeable. The conventional color television camera with its three spectral channels red, green and blue is therefore out of the question. The multi-channel system is achieved with the aid of the lighting unit 4 . With optical filter disks, only the light of the desired wavelengths is selected. The wavelength of the light generated can be changed by switching another filter into the beam path. In FIG. 4, a lighting unit is shown with two physical channels. It consists of a light source 15 . The light is preprocessed in the condenser 16 and passed into an optical waveguide 17 which splits into two channels. The filter disks mentioned sit in the filter supports 18 and 19 , which are each positioned with a stepper motor 20 and 21 . The light from the two channels is conducted in the fusion unit 26 into an optical waveguide 27 , which leads to the scanning unit. An aperture 23 is provided for switching between the two channels without inertia.

With this unit, any measuring wavelengths or spectral ranges represented by them are set will. It is therefore of central importance for that System. Are z. B. 6 measuring wavelengths required, recommended the use of a lighting unit at least 6 physical channels. Then the filter holder only set and need in the learning phase no longer switched to in the classification phase will.

A pen is divided into different regions. Homogeneous Regions are solid. Transition regions are regions the transitions of two or more homogeneous regions contain. Texture regions consist of multi-colored prints. A special case of a texture region is the embossing region. This is a single color print. Marking regions are therefore easier to analyze than that general texture regions. For each region in the Learning phase features determined and in a data structure saved. These characteristics are said to affect the regions concerned characterize so precisely that with their help, possibly existing defects can be recognized.

There are special features for each region type in the learning phase to determine. The measurement wavelengths are the basis of all characteristics. These measuring wavelengths are determined by for each Spectral characteristic pair of occurring colors the difference function is determined. The wavelengths at which Extremata occur are potential measuring wavelengths. These The quantity is reduced to one with the help of an effectiveness criterion minimal number reduced.

One of the main problems is the limited resolution the scanning matrix. Among other things, this enforces the  Introduction of transition regions. Because of the non-punctiform Scanning spectral characteristics arise that none actually represent existing color, but only a pseudo mixed color that occurs because the measuring surface several sub-areas of different primary colors contains. It is for determining the measuring wavelengths however, the spectral characteristics of these primary colors are required to know. For homogeneous regions with embossed regions this is easy. In the case of general texture regions offers a procedure that first with help the amount of spectral vectors measured in a cluster analysis reduced by spectral vectors to cluster areas be summarized. After that, the convex hull becomes this reduced vector set and linearly interpolated them. The corner vectors of this smoothed convex hull are the spectral vectors of the primary colors sought. These Analysis is only carried out in the learning phase.

While there is enough time for the learning phase (e.g. 30 seconds per pen), must be due to the required Lead time the classification of the pens in parallel work be carried out, what only by using special Hardware can be realized. The pipeline principle provides a method by which in different processing units processed several pens simultaneously can be, the data of these pens in different Phases. An essential feature is a decentralized storage.

Overall, the classification task is broken down into a large number of small sub-tasks. Each subtask is processed by an independently operating computing unit with its own processor, which is specially tailored to this task. So there are z. B. a separate processing unit for each region type. The homogeneous regions of a pen can be examined simultaneously (see Fig. 21). If necessary, a separate sub-processing unit is provided for each transition region and each embossing region of a pen. The assignment of the individual sub-processing unit to the regions and their initialization takes place in the learning phase. The process control is carried out with a decentralized control.

Claims (25)

1. Device for checking products for surface manufacturing defects with regard to color, texture, embossing or the like, in particular for checking surfaces of mine pencils, consisting of a transport unit, a lighting unit, a positioning unit, a scanning unit, a control unit and evaluation and an ejection unit, characterized in that

• a) the transport unit ( 1 ) transports the test objects in the correct position in the test zone,
• b) the illumination unit ( 4 ) acts on the surfaces to be tested in a temporal sequence and / or at the same time with radiation of certain fixed, predetermined different wavelength ranges,
• c) the scanning unit ( 2 ) measures the light intensity for each pixel and each wavelength range and stores it in a memory location,
• d) the evaluation unit ( 7 ) compares the detected values with reference values and makes a "good-bad" decision, which is implemented in the ejection unit ( 6 ).

2. Device according to claim 1, characterized in that the lighting unit ( 2 ) acts on the test object surface in chronological order with automatically adjustable test wavelengths. 3. Apparatus according to claim 1 or 2, characterized in that the test specimens in the scanning unit ( 2 ) are rotated such that the entire test object surface can be scanned in succession. 4. Device according to one of claims 1 to 3, characterized in that in the evaluation unit ( 7 ) a comparison between the measured value and the reference value is carried out. 5. Device according to one of claims 1 to 4, characterized characterized that using predetermined criteria the position of the test object is determined. 6. Device according to one of claims 1 to 5, characterized characterized in that the position of the scanning grid with respect is determined on the reference object. 7. The device according to claim 6, characterized in that through constant comparison of target and actual values, a fine adjustment he follows. 8. Device according to one of claims 1 to 7, characterized characterized that one on certain types of regions (homogeneous regions, transition regions, embossing regions, Texture regions) specialized evaluation subunit is available in the with in the reference memory stored data of this region taking into account a predefined criteria a "good-bad" decision is hit. 9. The device according to claim 8, characterized in that a corresponding time-delayed signal is given to the ejection unit ( 6 ) by linking the results of the evaluation subunit. 10. Device according to one of claims 1 to 9, characterized characterized in that in the case of a "good" decision  the reference values with the help of moving, averaging Operations are updated. 11. Device according to one of claims 1 to 10, characterized in that the lighting unit ( 4 ) has mechanically switchable filter carrier ( 8 ) and the radiation is transported with the aid of optical fibers or bundles ( 5 ). 12. The apparatus according to claim 11, characterized in that the beam path ( 17 ) coming from the radiation source ( 15 ) is split into a plurality of parallel beam paths, that in each of the parallel beam paths a movably mounted filter carrier ( 18, 19 ) is introduced which can be positioned electromechanically and thus allows a selectable portion of the radiation from the radiation source to pass through and that the parallel beam paths ( 24, 25 ) can be combined into a single beam path ( 27 ) with the aid of an optical switch ( 26 ). 13. The apparatus according to claim 12, characterized in that in the parallel beam paths ( 24, 25 ) there is an electromechanical or electro-optical switch ( 25 ) which serves as a light barrier and releases one or more beam paths and the light flow in all others Beam paths blocked. 14. Device according to one of claims 1 to 13, characterized in that the scanning unit ( 2 ) for parallel scanning essentially has the shape of the test objects ( 32 ). 15. Device according to one of claims 1 to 14, characterized in that both scanning and lighting elements ( 30 ) are integrated in the scanning unit ( 2 ). 16. The device according to one of claims 1 to 15, characterized characterized in that charge-coupled for sampling Semiconductor matrices are provided, the from Test object reflected radiation with the help of lens optics is focused on the semiconductor matrix. 17. The device according to one of claims 1 to 16, characterized in that the scanning units ( 42 ) are surrounded by optical fibers ( 41 ) of the lighting unit for homogeneous illumination of the surface of the measurement object. 18. Device according to one of claims 1 to 17, that the scanning unit ( 42 ) consists of several sub-units which are arranged in series. 19. Device according to one of claims 1 to 18, characterized characterized in that a preprocessing unit is present, which serves to correct the samples and processed the measured values of the entire scanning unit and corrected that by scanning a uniform Screening of the test object area is carried out and off overlapping measuring points of adjacent scanning units one measuring point each is formed and stored becomes. 20. Device according to one of claims 1 to 19, characterized in that one of a control unit ( 56 ), a counter ( 54 ), a plurality of comparison units ( 53 ), a pattern memory ( 51 ), a reference memory ( 55 ) and one Protocol memory ( 52 ) existing processing unit for the analysis of homogeneous regions is present, the counter ( 54 ) addressing all memory cells of the sample memory ( 51 ) in sequence, activating a memory line of the reference memory ( 55 ) and the associated memory line of the sample memory ( 51 ) at the same time If a comparison unit ( 53 ) for each memory column checks the value of the addressed memory cell of the sample memory ( 51 ) to determine whether it lies within an interval specified by the relevant memory cell of the reference memory ( 55 ), information in the log memory ( 52 ) relating to the Values of the sample memory ( 51 ) that are not in the vo Given interval and that finally the control unit only delivers a "good" message to the unit to be called when all the memory cells of the sample memory ( 51 ) to be checked are within the predetermined value interval. 21. Device according to one of claims 1 to 20, characterized characterized in that a control unit, arithmetic unit, Counter, sample memory, buffer and Existing evaluation unit for analysis of transition regions is present, the counter addresses the lines of the pattern memory in order and the sample lines in the direction of the transition transition run, the computing units, each are assigned to a column of the sample memory, from the value of the last addressed memory cell and the value of the currently addressed memory cell Determine and update characteristic values as well as in the buffer record, these parameters provide information give about the intensities of the measured values at the beginning of the Transition region and at its end as well as over the Transition change speed and location of the transition center, with the buffer after Counting up the counter is read and the characteristic values  each column with those of the previous column as well compared with reference values according to specified criteria become and every time one is exceeded Limit specified deviation the position of the measuring line is noted in the scanning grid and finally only then delivered a "good" signal from the control unit is recognized when each pattern column is good and with the characteristic values of the sample columns recognized as good immediately a correction of the reference values is calculated becomes. 22. The device according to one of claims 1 to 22, characterized characterized that an evaluation unit for analysis transition regions of three homogeneous regions is, the measurement data of the potential transition region including the reference values of the first Transaction a subunit are supplied, the measurement data the potential transition region including the reference values of the second transaction of another Subunit to be fed and finally by evaluating the log memory of the two subunits formed an overall "good / bad" statement the transaction region is then error-free, if the intersection of those in each subunit is considered faulty recognized sub-regions delivers the empty quantity. 23. The device according to one of claims 1 to 22, characterized characterized in that a positioning unit for the test specimen is present, their receiving bed of the shape the test specimen is adapted and in which the test specimen held by suction air and ejected by compressed air will. 24. The device according to one of claims 1 to 23, characterized characterized in that each computing unit of the first  Processing level is assigned to an elementary region and each arithmetic unit on the memory cell of the sample memory can access. 25. Device according to one of claims 1 to 24, characterized characterized in that the processing unit is a parallel processing the samples of columns of the sample memory carries out and the process when a faulty elementary region.