CN103440348A - Vector-quantization-based overall and local color image searching method - Google Patents

Abstract

The invention relates to a vector-quantization-based overall and local color image searching method. The invention provides a new color image searching method, and relates to the field of an image processing technology. The method comprises the steps of converting an RGB (red, green and Blue) color space into an HSV (Hue, Saturation, Value) color space, performing relatively accurate clustering division on the color space by applying a neural-network-based competitive learning algorithm training code book; describing a space distribution situation of colors by introducing a color transfer matrix; combining the two characteristics, namely an index column diagram and a main color transfer matrix, so as to perform the similarity measurement; processing images by applying a morphological opening-and-closing operation, highlighting a target contour so as to extract a local interest area so as to highlight an important area and limit background information. The color image searching method overcomes the defects that the color space distribution description is not enough and the background information cannot be effectively limited by using the overall color histogram method. By using the vector-quantization-based overall and local color image searching method, the color quantization is relatively accurate, the matching effect is relatively good, and the vector-quantization-based overall and local color image searching method is an effective method for further improving the researching efficiency.

Description

A kind of overall situation and local color-image retrieval method based on vector quantization

Technical field

The invention belongs to the CBIR field, be specifically related to the color search method that a kind of overall situation based on vector quantization and local region of interest combine.

Background technology

Along with the development of computer technology, multimedia technology and network technology, the great amount of images data are by the internet wide-scale distribution.Yet effective image search method for want of, make utilization for huge image data base always in extremely low efficiency.Search method for view data has three kinds usually: freely browse, text based image retrieval (Text Based Image Retrieval, TBIR) and CBIR (Content Based Image Retrieval, CBIR).Freely browsing and be only applicable to cas fortuit, is inappropriate for the professional client of the special multimedia messages of frequent use.There is the problem of two aspects in the text based image retrieval: the one, and require a great deal of time to each width view data is carried out manual word note and classified in order to building storehouse; The 2nd, this note often can't precision for the expression of personalized mankind's subjective vision content.In the CBIR system, image is to replace key word by the vision content for self to explain acquisition characteristic information, for example color, texture and shape information.They more approach the mankind's vision system.Image retrieval technologies is generally applied the low-level image feature of color-based at present, based on morphological images, processes, based on technology such as vector quantizations.

Color is the coloured image bottom, physical features the most intuitively, usually the variation of degeneration, size, resolution and the direction etc. of noise, picture quality being had to very strong robustness, is one of feature of using in most CBIR multimedia databases.Color characteristic commonly used mainly comprises color histogram, colour consistency vector, color correlogram and color matrix etc.Wherein, utilizing the color histogram retrieval is the most basic method, but lacks the description to color space information; The color correlogram is emphasized the space length correlativity of same color in image; Color matrix is mainly to adopt average and the variance of each color in image to compare, and processes simply, can use it as the initial survey of image retrieval, for hunting zone is dwindled in next step retrieval.Fig. 1 is a kind of image search method process flow diagram based on global color.

Mathematical morphology is one and take algebra of sets as basis, a new branch of science of research digitized video morphosis and fast parallel processing.The basic thought that morphological images is processed is to utilize an information that is called " probe " collection image of structural element.When probe constantly moves in image, just can consider the mutual relationship between the image various piece, thereby understand the architectural feature of image.Structural element is most important, the most basic concept during morphology is processed, and its effect in morphological transformation is equivalent to " filter window " in the signal processing.For same width image, the structural element difference, the result of processing is also different.In bianry image morphology application, the selection principle of structural element often has rotational invariance, or mirror invariant performance at least, that is to say, the initial point of structural element is at its geometric center place, and other pixels are symmetry shape about this initial point.The structural elements of commonly using have that level is single-row, vertically single-row, cross, disk, rhombus and square etc.Constantly perfect along with Mathematical Morphology theory, it is more and more extensive that mathematical morphology is cut apart middle application at image.

In recent years, the image search method based on vector quantization becomes the popular domain of many scholar's research.Block of image pixels is quantized, realized transmission and the coupling of image by the index of transmission or coupling code word.The process quantized can be regarded one as from k dimension space R ^kto the mapping of one of them finite subset Y, Q:R ^k→ Y={Y ₁, Y ₂..., Y _n.Its ultimate principle is k dimension space R ^kexhaustively be divided into N mutually disjoint subspace (cell) R ₁, R ₂..., R _n.At each subspace R _iin find out a representative vector Y _i={ y _i1, y _i2..., y _ik, be designated as vector set Y={Y ₁, Y ₂..., Y _n, Y is called code book or code book, Y _ibe called code word or code vector, N is the code book size.The vector quantization process is exactly to an input vector X={x ₁, x ₂..., x _k, find out a Y the most close with X in Y _ireplace X, i.e. Y _iit is the quantized value of X.

In the situation that given image to be retrieved, the image how to find the user to want from image library quickly and accurately is the problem that all kinds of image retrieval technologies are researched and solved, and the performance of improving image indexing system has become the previous considerable research topic of order.But the search method color space of prior art quantizes out of true, the space distribution of color describes not enough and local important information is not outstanding, is difficult to obtain the information of accurate retrieval.

Summary of the invention

The present invention is directed to color space in existing color-image retrieval method and quantize out of true, the space distribution description deficiency of color and local important information be distinct issues not, propose the color-image retrieval method.

The technical scheme that the present invention solves the problems of the technologies described above is, a kind of overall situation and local color-image retrieval method based on vector quantization, comprise step: read color image data, it,, from the RGB color space conversion to the hsv color space, is chosen to 4 * 4 pixels of adjacent and non-overlapping copies as trained vector; Adopt the method for trained vector quadratic sum sequence to form inceptive code book; Inceptive code book is practiced in the image construction trained vector training of choosing in image library; By adding up percentage frequency that each color occurs and the change color situation of adjacent pixel blocks, form color index histogram and main color transition matrix, using it as retrieval character; Utilize morphological images to process, highlight objective contour to extract topography interested zone; Utilize global color feature and local region of interest color characteristic Weighted Searching.

Specifically comprise:

Color space quantizes: the image that (1) is chosen a width rich color and is evenly distributed, and from the RGB color space conversion to the hsv color space, and by its H, S, tri-components of V extract; (2) select 4 * 4 pixels of adjacent and non-overlapping copies as trained vector; (3) adopt the method for vector quadratic sum sequence to form inceptive code book; (4) choose image as the training plan image set in all kinds of images of image library, with Competitive Learning Algorithm (CL algorithm) training inceptive code book.(5), through color cluster, three-dimensional color image can obtain respectively three code book H _i, S _i, V _i, they are merged into to an eigenvector (code book) ω, ω=(H _i, S _i, V _i) i=1,2 ..., N, and the set of codewords indexes is equivalent to a color look up table that comprises N kind color.

Color characteristic extracts: (1) color index histogram: according to color look up table, image in image to be retrieved and image library is divided into to 4 * 4 block of pixels, each block of pixels characterizes by the call number (being also a kind of color) of a code word; By adding up frequency and the percentage of each codewords indexes appearance, obtain the color index histogram H (v of coloured image ₁, v ₂, v _i... v _n); Wherein, v _imean the ratio of the code word appearance that call number is i, N is the code book size.(2) main color transition matrix: image is divided into to m * n piece, and each piece all comprises s * t pixel; Draw the main color index value (also i.e. this piece in the maximum index value of occurrence number) of each piece, form the main color matrix of a two dimension, its size is m * n, is designated as A={a _i,ji=1,2 ... m, j=1,2 ... n; Set up the matrix P of a N * N, the initial value of its each element is 0; Matrix A is scanned by Z-shaped, established a _i,jand a _p,qfor a pair of color (a in succession occurred in scanning sequence _i,ja in front _p,q), respective element in P

from increasing 1, so repeatedly, until scanned; The ratio occurred all block of pixels centerings of entire image of the neighbouring relations of variant color and this relation in statistical picture.

Region of interesting extraction: adopt structural element tolerance, extract the correspondingly-shaped in image, reduced data, keep the basic configuration feature and remove incoherent structure; Utilize opening operation to eliminate loose point and the burr less than structural element, cut off elongated overlap joint and play the effect of separation, image is carried out to level and smooth and low-pass filtering; Utilize closed operation that the breach less than structural element or hole are filled above, overlap short interruption object is coupled together, image is carried out to filtered external, polish the protruding wedge angle to image inside; Through opening and closing operation, image is carried out to smoothing processing, retain the important profile of image, remove the details and the noise that easily cause undue division; Then set corresponding threshold value and eliminate tiny piecemeal, take out the piece of area maximum; The ranks value of record start pixel and width and height are to extract the rectangle area-of-interest.

Similarity coupling: calculate Image Visual Feature similarity in image to be checked and image library, by the overall situation and the retrieval of local region of interest characteristic weighing.If image to be checked is A, in image library, any piece image is B, and their area-of-interest is respectively a and b, and its corresponding color index histogram and main color transition matrix are respectively: H _a, H _b, H _a, H _b, D _a, D _b, D _a, D _b; Calculate overall similarity

$\mathrm{Simi}1={\mathrm{\ω}}_1\sqrt{{(H_A-H_B)(H_A-H_B)}^T}+{\mathrm{\ω}}_2\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{A_{i,j}}-D_{B_{i,j}})}^2}$ And local similarity

$\mathrm{Simi}2={\mathrm{\ω}}_3\sqrt{{(H_a-H_b)(H_a-H_b)}^T}+{\mathrm{\ω}}_4\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{a_{i,j}}-D_{b_{i,j}})}^2},$ ω ₁for the weight of global color histogram, ω ₂for the weight of the main color transition matrix of the overall situation, ω ₃for the histogrammic weight of local color index, ω ₄for the weight of the main color transition matrix in part (ω wherein ₁, ω ₂, ω ₃, ω ₄∈ [0,1], ω ₁+ ω ₂=1, ω ₃+ ω ₄=1); Overall similarity and local similarity is synthetic as final tolerance: Similar=pSimi1+qSimi2(is p wherein, q ∈ [0,1] p+q=1); Similar is arranged by ascending order, and the image that the similarity value is less is more similar to query image; Return on demand the result for retrieval after sequence.

The present invention adopts the method based on the sequence of vector quadratic sum to select inceptive code book, uses the Competitive Learning Algorithm training code book based on neural network, and color space is carried out to more accurate clustering, makes quantized result more approach people's perception; Introduce the color transition matrix on the basis of extracting piecemeal master color, to describe the space distribution situation of color; Index histogram and two kinds of color characteristics of main color transition matrix are combined and carry out similarity measurement; Use the morphology opening and closing operation to process image, highlight objective contour, to extract local region of interest; Compose with different weights to image overall and local region of interest, both reflect to a certain extent the space distribution situation of color of image, given prominence to again important area, the restriction background information.This programme has overcome the global color histogram method color space is distributed and describes not, can not effectively limit the shortcoming of background information, the effective cluster of Competitive Learning Algorithm similar vector, construct robustness code book preferably, color space is quantized more accurate, matching effect is better, is the effective ways that further improve recall precision.

The accompanying drawing explanation

Fig. 1 is a kind of image search method process flow diagram based on global color;

The basic flow sheet that Fig. 2 is the inventive method;

Fig. 3 is Codebook Design (color quantizing) algorithm flow chart;

Fig. 4 utilizes morphological images to process the algorithm flow chart that extracts area-of-interest;

Fig. 5 is whole precision ratio curve map.

Embodiment

Hereinafter the meaning of the variable that uses is as follows: X means trained vector; Y means code book; ω means the final code book that three color components are synthetic; N means the code book size; H means the color index histogram; D means main color transition matrix; P means precision ratio.

The present invention quantizes and local region of interest from color space, improves the color quantizing precision, fully reflects the color of image distribution situation and strengthens image local feature to improve retrieval performance.The invention will be further described below to use concrete example and accompanying drawing, the basic flow sheet that Fig. 2 is the inventive method.Concrete implementation step is as follows:

Color space quantizes: choosing the HSV space is the color quantizing space, carries out the conversion of rgb space to the HSV space; By image H, S, the V space extracts, and selects 4 * 4 pixels of adjacent and non-overlapping copies as trained vector; Adopt the method for vector quadratic sum sequence to form inceptive code book; And form final code book (vector quantization code table) with CL Algorithm for Training inceptive code book, form a color look up table (color set).

Color characteristic extracts: color characteristic is selected color index histogram and main color transition matrix.

(1) color index histogram: the color of each small pixel piece (4 * 4) of image in image to be retrieved and image library is summed up as to its corresponding codewords indexes value according to color look up table; By adding up frequency and the percentage of each codewords indexes appearance, obtain the Color Statistical vector of coloured image; The call number of each code word of take is horizontal ordinate, and the ratio of its appearance is ordinate, obtains the color index histogram of image.

(2) main color transition matrix: by image block (every small pixel piece that comprises t 4 * 4), according to color look up table, add up index value that in each piece, color frequency the is maximum main color as this block of pixels; Scan each block of pixels by Z-shaped, the main color value that statistics occurs in succession; The ratio that in document image, the neighbouring relations of variant color and this neighbouring relations occur all block of pixels centerings of entire image.

Region of interesting extraction: utilize morphologic opening and closing operation to carry out smoothing processing to image, retain the important profile of image, remove the details and the noise that easily cause undue division; Then set corresponding threshold value and eliminate tiny piecemeal, take out the piece of area maximum; The ranks value of record start pixel and width and height are to form the rectangle area-of-interest.

Similarity coupling: the Euclidean distance of calculating respectively the overall situation and local color index histogram and main color transition matrix; Euclidean distance weighting by global color index histogram and main color transition matrix, be overall similarity; Euclidean distance weighting by local color index histogram and main color transition matrix, be local similarity; The synthetic overall situation and local similarity are also arranged by ascending order.

The following specifically describes implementation of the present invention:

(1) inceptive code book design: the image of choosing a width rich color and being evenly distributed, by its H, S, the V space extracts, and selects 4 * 4 pixels of adjacent and non-overlapping copies as trained vector; Adopt the method for vector quadratic sum sequence to form inceptive code book.Fig. 3 is Codebook Design (color quantizing) algorithm flow chart.The concrete steps of vector quadratic sum ranking method are:

1. establishing trained vector integrates as X={X ₁, X ₂..., X _l, code book is of a size of N;

2. calculate the X of each vector _lquadratic sum S _l, by S _lby ascending order, arrange;

3. the trained vector after sequence is divided into to the N section, every section has K=L/N trained vector;

4. select successively first code word of every section as initial code word, form the inceptive code book that is of a size of N.

This inceptive code book algorithm uses the method segmentation based on the sequence of vector quadratic sum to choose initial code word to trained vector.Algorithm has been used the characteristic quantity of vector, has broken away from the dependence to the picture structure factor, forms robustness inceptive code book preferably.

(2) code book training: choose 24 totally different width images of color as the training plan image set in all kinds of images of image library, and train inceptive code book with Competitive Learning Algorithm (CL algorithm).

Competitive Learning Algorithm (CL algorithm) concrete steps:

1. the trained vector of establishing piece image integrates as X={X ₁, X ₂..., X _land X _l∈ X, code book to be designed is Y={Y ₁, Y ₂..., Y _n, iterations is t, through the inceptive code book design, obtains N initial codebook

and adopt square error to estimate;

2. the error between calculation training vector and each code word is estimated

$d_j^{\left(t\right)}=(X_l,Y_j^{(t-1)})={\left|\right|X_l-Y_j^{(t-1)}\left|\right|}^2,j=\mathrm{1,2},\·\·\·,N;$

3. select least error to estimate corresponding code word Y _i, the code word that current competition is won,

d _i=min(d _j),j=1,2,…,N；

4. press following formula adjustment triumph unit code word:

Wherein, a ^(t)for learning rate, get a here ^(t)=1/t;

5. deconditioning when meeting error requirements or given number of iterations, gained Y is as final code book; Otherwise, go to step 2..

The competitive learning vector quantization is a kind of simple hard decision clustering algorithm, only upgrades the neuron (code word) of winning in the process of study, and continuous regularized learning algorithm speed, makes algorithm convergence.Through color cluster, three-dimensional color image can obtain respectively three code book H _i, S _i, V _i, they are merged into to an eigenvector (code book) ω, ω=(H _i, S _i, V _i) i=1,2 ..., N, and the set of codewords indexes is equivalent to a color look up table that comprises N kind color.

(3) color index histogram: in the color histogram based on pixel, for piece image I, its color has N level (C ₁, C ₂..., C _n) form C _iit is i level color value.In entire image, there is C _ithe number of pixels of value is ‖ L _ci‖, the statistical value h of one group of pixel ₁, h ₂..., h _njust become the color histogram of this image.It is defined as: H (I)=(h ₁, h ₂..., h _n) wherein, h _i=‖ L _ci‖/m; C _ithe feature value of representative image; But N is the number of feature value; ‖ L _cihaving color feature value in the ‖ presentation video is C _inumber of pixels; M is the total number of pixels of image.

At first index histogram based on vector quantization is divided into image the block of pixels identical with codeword size, then each piece is carried out to vector quantization, output quantization manipulative indexing, then computation index sequence histogram (quantity of bin equals the size of code book).Wherein the quantity of block of pixels is much smaller than pixel quantity, and calculated amount can effectively reduce.When two block of pixels have same or analogous pixel histogram, its content is obviously different, with respect in pixel processing method, being considered as identical block of pixels, in vector quantization, they generally can not be quantified as same code word, are also corresponding different index values.The histogrammic concrete formation step of index is as follows:

1. image in image to be retrieved and image library is divided into to 4 * 4 block of pixels;

2. according to color look up table, each block of pixels characterizes by the call number (being also a kind of color) of a code word;

Concrete grammar is described below:

1) calculate the Euclidean distance between each trained vector and each code word;

2), for each trained vector, the codewords indexes of selection and its Euclidean distance minimum is as the sign of this vector;

3) record the codewords indexes value that all vectors are corresponding (being characterization).

3. by frequency and the percentage of each codewords indexes appearance of statistics, obtain the color index histogram H (v of every width coloured image ₁, v ₂, v _i... v _n); Wherein, v _imean the ratio of the code word appearance that call number is i, N is the code book size.

(4) main color transition matrix: main color transition matrix be based upon on HSV quantized color space in order to portray a kind of color characteristic of the relative distributing position of color.At first by image block, according to every color frequency of color look up table statistics, maximum index value is as the main color of this block of pixels; Scan each block of pixels by Z-shaped, the main color value that statistics occurs in succession; The ratio that in document image, the neighbouring relations of variant color and this neighbouring relations occur all block of pixels centerings of entire image.The concrete grammar that obtains image master color transition matrix is described below:

1. obtained the vector quantization code table of image by the CL Algorithm for Training, N code word, also be quantized into N kind color altogether;

2. image is divided into to m * n piece, each piece all comprises s * t pixel;

3. draw the main color index value of each piece, namely the maximum index value of this piece occurrence number.So just formed the main color matrix of a two dimension, its size is m * n, is designated as

A={a _i,j}i=1,2,…m,j=1,2,…n；

4. set up the matrix P of a N * N, the initial value of each element is 0.Matrix A is scanned by Z-shaped, established a _i,jand a _p,qfor a pair of color (a in succession occurred in scanning sequence _i,ja in front _p,q) respective element in P

from increasing 1, so repeatedly, until scanned;

5. set up the matrix D of a N * N, the computing formula of its element is as follows

(5) region of interesting extraction: its basic thought is to choose suitable structural element according to the target signature of original image, utilize structural element original image to be carried out to the computings such as translation, intersecting and merging, smoothed image also highlights objective contour, then important contour area is formed to a rectangular area, output coordinate point.At first utilize opening and closing operation to carry out smoothing processing to image, retain the important profile of image, remove the details and the noise that easily cause undue division; Then set corresponding threshold value and eliminate tiny piecemeal, take out the piece of area maximum; The ranks value of record start pixel and width and height are to form the rectangle area-of-interest.Fig. 4 utilizes morphological images to process the algorithm flow chart that extracts area-of-interest.

The fundamental operation that morphological images is processed has expansion, burn into opening operation and closed operation.

Expand: mathematical definition is set operation, and A is expanded by B, is designated as A ⊕ B, is defined as

Wherein,

for empty set, B is structural element.It is the set that all structural element origin positions form that A is expanded by B, and the B after mapping translation is at least overlapping with some part of A.Expansion is that the institute contacted with object is had powerful connections and merges to this object, makes the process of border to the outside expansion, can be used for filling up the cavity in object.Concrete steps are as follows:

1. use each pixel of structural element B scan image A;

2. do AND-operation with the bianry image of structural element and its covering;

If be all 3. 0, this pixel of result images is 0, otherwise is 1.

Corrosion: A is corroded by B, is designated as A Θ B, is defined as

Wherein,

for empty set, B is structural element.A is the set that all structural element origin positions form by the B corrosion, and the B after mapping translation and the background of A do not superpose.Corrosion is a kind of elimination Debris, makes the process of border to internal contraction, can be used for eliminating little and insignificant object.Concrete steps are as follows:

1. use each pixel of structural element B scan image A;

2. do AND-operation with the bianry image of structural element and its covering;

If be all 3. 1, this pixel of result images is 1, otherwise is 0.

Opening operation: first corrode afterwards and expand, A carrys out the result after dilation erosion with B after being corroded by B again, is defined as:

Wherein, ∪ { } refers to all union of sets collection in braces, symbol

mean that C is the subset of D.The simple geometry of this formula is interpreted as: A o B is B union of translation of coupling fully in A.Opening operation has been deleted the subject area that can not comprise structural element fully, and level and smooth object outline, disconnected narrow connection, removed tiny outshot, simultaneously and its area of not obvious change.

Closed operation: the post-etching that first expands, A corrodes the result after expansion with B after being expanded by B again, is defined as:

Wherein, ∪ { } refers to all union of sets collection in braces, from geometrically say A B be all not with the translation union of the overlapping B of A.As opening operation, closed operation meeting smooth object profile.Yet different from opening operation, closed operation generally can couple together narrow breach and form elongated curved mouth, and the little hole of packing ratio structural element, while its area of not obvious change.

The concrete steps of extracting area-of-interest are as follows:

1. read coloured image, be converted into gray level image;

2. gray level image is carried out to filtering, smoothed image is removed noise;

3. choose the disc structure element, gray level image is carried out to opening operation;

4. the image after opening operation is subtracted to computing, to strengthen image, eliminated background;

5. the gray level image after strengthening is carried out to contrast stretching, gray-scale value is mapped to [0,1];

6. by Binary Sketch of Grey Scale Image;

7. choose the rectangular configuration element, respectively the rectangle row, column is carried out to opening and closing operation;

8. mark is carried out in the zone be communicated with;

9. the characteristic dimension of each connected region of computed image, take out the zone of area maximum, records the ranks value of its initial pixel and height and width with the formation rectangular area;

10. according to the large young pathbreaker rectangular area intercepting of code word, be suitable size.

(6) similarity coupling: for an image to be retrieved, at first extract its area-of-interest, record the ranks coordinate of its starting and ending pixel, in corresponding intercepting image library, the respective regions of each image carries out the color characteristic coupling as area-of-interest.Finally by global color feature and the retrieval of local region of interest characteristic weighing.Concrete steps are as follows: establishing image to be checked is A, and in image library, any piece image is B, extracts its area-of-interest, is respectively a and b;

1. the percentage frequency of the statistics overall situation and local each color, form color histogram H _a, H _b, H _a, H _b;

2. add up the change color situation of adjacent pixel blocks, form color transition matrix D _a, D _b, D _a, D _b;

3. adopt Euclidean distance to calculate overall similarity:

ω wherein ₁for the histogrammic weight of global color index, ω ₂weight (ω for the main color transition matrix of the overall situation ₁, ω ₂∈ [0,1] and ω ₁+ ω ₂=1),

for color transition matrix D _a, D _bin the capable j column element of i;

4. adopt Euclidean distance to calculate local similarity:

ω wherein ₃for the histogrammic weight of local color index, ω ₄weight (ω for the main color transition matrix in part ₃, ω ₄∈ [0,1] and ω ₃+ ω ₄=1),

for color transition matrix, D _a, D _bin the capable j column element of i;

5. synthesize similarity: Similar=pSimi1+qSimi2(is p wherein, q ∈ [0,1] p+q=1);

6. Similar is arranged by ascending order, return to result for retrieval.

Under the MATLAB7.9 software platform, the example to the present invention program is elaborated by reference to the accompanying drawings

Use the color image data source of 256 * 384 specifications, compare by emulation experiment and a kind of image retrieval algorithm based on global color.The present invention program's concrete implementation step is as follows:

The inceptive code book design phase:

Step 1: the coloured image that reads a width rich color and be evenly distributed obtains 3 of color image data source and ties up matrix A (256 row are arranged, and 384 are listed as, and 3 Color Channels, note by abridging as (256 * 384 * 3));

Step 2: be the hsv color space by the RGB color space conversion, extract respectively the H(tone), S(saturation degree), V(brightness) three components.Wherein Parameter H means color information, i.e. the position of spectral color of living in, and it means with angular metric, red, green, blue 120 degree of being separated by respectively, complementary colors differs respectively 180 degree.Saturation degree S is a ratio value, scope from 0 to 1, and it means the saturation degree of selected color and the ratio between this color maximum saturation.V means the bright degree of color, scope from 0 to 1;

Step 3: adopt the method for vector quadratic sum sequence that H, S, its size of V(are to 256 * 384) three element quantizations become three inceptive code books.Take tone H as example, establish code book and be of a size of N, at first convert 4 * 4 block of pixels of adjacent non-overlapping copies to 1 * 16 row vector, entire image is converted to i.e. 6144 row vectors of 6144 * 16(by 256 * 384); Calculate the quadratic sum S of 6144 vectors _l, by S _lby ascending order, arranged; The trained vector that will sort is divided into the N section, and every section has T=6144/N trained vector; Select successively first code word of every section as initial code word, forming size N is inceptive code book.

The code book training stage:

Step 1: choose 24 totally different width images of color as the training plan image set in all kinds of images of image library;

Step 2: the trained vector of establishing piece image integrates as X={X ₁, X ₂..., X ₆₁₄₄and X _l∈ X, code book to be designed is Y={Y ₁, Y ₂..., Y _n, iterations is t, through the inceptive code book design, obtains N initial codebook

and adopt square error to estimate;

Step 3: the error between calculation training vector and each code word is estimated

$d_j^{\left(t\right)}=(X_l,Y_j^{(t-1)})={\left|\right|X_l-Y_j^{(t-1)}\left|\right|}^2,j=\mathrm{1,2},\·\·\·,N;$

Step 4: select least error to estimate corresponding code word Y _i, the code word that current competition is won,

d _i=min(d _j),j=1,2,…,N；

Step 5: press following formula adjustment triumph unit code word:

Wherein, a ^(t)for learning rate, get a here ^(t)=1/t;

Step 6: deconditioning when meeting error requirements or given number of iterations, gained Y is as final code book; Otherwise, go to step 3;

Step 7: successively the image of training plan image set is trained.

After CL algorithm color cluster, three-dimensional color image can obtain respectively three code book H _i, S _i, V _i(i=1,2 ..., N), three color components after quantizing are merged into to an eigenvector (code book) ω, ω=(H _i, S _i, V _i)=L _hh _i+ L _ss _i+ L _vv _i, and the set of codewords indexes is equivalent to a color look up table that comprises N kind color.

The color index histogram:

Step 1: in image just to be retrieved and image library, image is divided into the block of pixels (as be 4 * 4 in this test) with the code word formed objects;

Step 2: according to color look up table, each block of pixels characterizes by the call number (being also a kind of color) of a code word, and specific implementation method is as follows:

4) calculate the Euclidean distance between each trained vector and each code word;

5), for each trained vector, the codewords indexes of selection and its Euclidean distance minimum is as the sign of this vector;

6) record 6144 codewords indexes values (being characterization) that vector is corresponding.

Step 3: by adding up the percentage frequency of each index appearance, obtain the color index histogram H (v of every width coloured image ₁, v ₂, v _i... v _n); Wherein, v _imean the ratio of the code word appearance that call number is i, N is the code book size.

Main color transition matrix:

Step 1: quantized to obtain the vector quantization code table of image by the color space of front, N code word, also be quantized into N kind color altogether;

Step 2: image is divided into to m * n piece, and each piece all comprises this experiment of s * t(and chooses 4 * 4) individual pixel;

Step 3: draw the main color index value of each piece, namely the maximum index value of this piece occurrence number.So just formed the main color matrix of a two dimension, its size is m * n, is designated as A={a _i,ji=1,2 ... m, j=1,2 ... n;

Step 4: set up the matrix P of a N * N, the initial value of each element is 0.Matrix A is scanned by Z-shaped, established a _i,jand a _p,qfor a pair of color (a in succession occurred in scanning sequence _i,ja in front _p,q) respective element in P

from increasing 1, so repeatedly, until scanned;

Step 5: set up the matrix D of a N * N, the computing formula of its element is as follows

d is exactly the main color transition matrix of this image.

Region of interesting extraction:

Step 1: read coloured image, and be converted into gray level image;

Step 2: gray level image is carried out to filtering, and from the input data, filtering noise and interference are to extract useful information;

Step 3: use IPT function strel to construct various shapes and size structure element (it is structural element that flat disc is chosen in this experiment);

Step 4: gray level image is carried out to opening operation, and deletion can not comprise the subject area of structural element, and level and smooth outline object, disconnect narrow connection and remove tiny connection, while its area of unconspicuous change;

Step 5: the image after filtered gray level image and opening operation is subtracted to computing, to strengthen image, eliminate background;

Step 6: the gray level image after strengthening is carried out to contrast stretching, threshold value m is set, by input value lower than in the gray level boil down to output image of m than the dark gray level than in close limit; Similarly, by input value higher than brighter gray level in the gray level boil down to output image of m than in close limit; The image that output has higher contrast, gray-scale value is mapped to [0,1];

Step 7: gray level image is chosen by suitable threshold value, obtain and still can reflect the binary image of integral image and local feature, what image was become is simple, and data volume reduces and highlight the objective contour of area-of-interest, to facilitate the further processing of image;

Step 8: choose the rectangular configuration element, respectively the rectangle row, column is carried out to opening and closing operation, the tiny noise distributed on tiny cavity, burr and background area in smooth object;

Step 9: mark is carried out in the zone be communicated with, generally adopt eight connections or four to be communicated with and find.If eight be communicated with refer to a pixel with other pixels in upper and lower, left and right, the upper left corner, the lower left corner, the upper right corner is being connected with the lower right corner, thinks that they are communicated with; Four are communicated with and refer to if the position of pixel upper and lower, the left or right adjacent in other pixels thinks that they are communicated with, in the upper left corner, the lower left corner, the upper right corner or the lower right corner connect, do not think their connections;

Step 10: calculate the characteristic dimension of connected region, take out the zone of area maximum, record the height in the ranks value of its initial pixel and zone and width to form rectangular area;

Step 11: the number of pixels mould of ranks 4 is removed to remainder, the rectangular area that forms the ranks value and be all 4 integral multiple is area-of-interest (because 4 * 4 sizes are got in the quantification of this experiment block of pixels, code word is 1 * 16, therefore rectangular area ranks pixel is required to be 4 integral multiple, to facilitate block of pixels, quantize).

The similarity coupling:

Step 1: establish two width image A and B, extract its area-of-interest, be respectively a and b;

Step 2: the percentage frequency of the statistics overall situation and local each color forms color index histogram H _a, H _b, H _a, H _b;

Step 3: the change color situation of statistics adjacent pixel blocks forms main color transition matrix D _a, D _b, D _a, D _b;

Step 4: adopt Euclidean distance as measure, calculate respectively the Euclidean distance of two width image overall index histograms and main color transition matrix, two kinds of color characteristics are weighted.Overall similarity is: $\mathrm{Simi}1={\mathrm{\ω}}_1\sqrt{{(H_A-H_B)(H_A-H_B)}^T}+{\mathrm{\ω}}_2\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{A_{i,j}}-D_{B_{i,j}})}^2}$ Wherein, ω ₁for the histogrammic weight of global color index, ω ₂weight (ω for the main color transition matrix of the overall situation ₁, ω ₂∈ [0,1] and ω ₁+ ω ₂=1);

Step 5: adopt Euclidean distance as measure, calculate respectively the Euclidean distance of two width image local index histograms and main color transition matrix, two kinds of color characteristics are weighted.Local similarity is: $\mathrm{Simi}2={\mathrm{\ω}}_3\sqrt{{(H_a-H_b)(H_a-H_b)}^T}+{\mathrm{\ω}}_4\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{a_{i,j}}-D_{b_{i,j}})}^2}$ Wherein, ω ₃for the histogrammic weight of local color index, ω ₄weight (ω for the main color transition matrix in part ₃, ω ₄∈ [0,1] and ω ₃+ ω ₄=1);

Step 6: synthetic similarity: Similar=pSimi1+qSimi2(is p wherein, q ∈ [0,1] p+q=1);

Step 7: Similar is arranged to (the less explanation two width images of Similar value are more similar) by ascending order;

Step 8: by demand, return to result for retrieval.

Be the correctness of result for retrieval for the evaluation of retrieval effectiveness, what mainly use is precision ratio (precision) and two indexs of recall ratio (recall).Precision ratio refers in the one query process, and system is returned to the number of associated picture in Query Result and accounted for all ratios of returning to picture number.Recall ratio refers to that system returns to the ratio that the number of associated picture in Query Result accounts for all associated picture numbers in image library (comprise return and do not return).The effect of these two higher explanation search methods of index is better.The formula of precision ratio and recall ratio can be expressed as:

Precision ratio $P=p\left(A\right|B)=\frac{p(A\∩B)}{p\left(B\right)}=\frac{R_A}{R_A+R_B}$

Recall ratio $R=p\left(B\right|A)=\frac{p(A\∩B)}{p\left(A\right)}=\frac{R_A}{R_A+R_C}$

Wherein, R _athe number of the associated picture that expression retrieves; R _bmean the unrelated images number retrieved; R _cundetected associated picture number in the presentation video storehouse.

The standard image data storehouse that experiment adopts Li.J to provide, choose wherein 420 256 * 384 or 384 * 256 coloured image forms the retrieving images storehouse, total personage, dinosaur, flower, grassland, seabeach, mountain peak and automobile seven class images, every class comprises 60 width, with precision ratio, weighs its retrieval performance.

During concrete test, different classes of image is respectively got 10 width and is formed the test pattern image set, and every width image returns to 6 width successively, 12 width ..., the retrieving images of 60 width.Return in the number situation in difference, calculate respectively the average precision of Different categories of samples image as such other precision ratio.Table 1 has provided the global color searching algorithm and algorithm of the present invention is 6 width returning to picture number, 12 width ..., the precision ratio of every class image in the situation of 60 width.

The precision ratio P(% of each class image of table 1)

Table 1 is continuous

As can be seen from Table 1, returning to picture number when less, the precision ratio of Global Algorithm and this algorithm is all higher, and along with the increase of returning to picture number, precision ratio all descends to some extent.The comparatively concentrated image for color distribution, as personage, dinosaur, flower etc., after adding local region of interest, when returning to the different images number, recall precision all is significantly improved.For seabeach, the relatively loose image of this two classes color distribution of mountain peak, after adding local region of interest, recall precision has some fluctuatings, as seabeach when returning to 30 width image, mountain peak is when returning to 48 width and 54 width, this paper recall precision descends to some extent, and this is because of the unconcentrated image of color distribution, and target and background are distinguished not obvious.For the precision ratio of image library integral body, Fig. 5 has provided accordingly result.

In the situation that it is identical to return to picture number, calculate the average precision of seven class images, the whole precision ratio as return to number hypograph storehouse at this, carry out comprehensive evaluation with this to algorithm.Because the image of each semantic classes has a great difference, everyone to the impression of image and understand also different, so precision ratio has certain fluctuation.From whole structure, the searching algorithm performance that the present invention proposes is more excellent than the global search algorithm.

Claims (5)

1. the overall situation and the local color-image retrieval method based on vector quantization, is characterized in that, reads color image data, and it,, from the RGB color space conversion to the hsv color space, is chosen to 4 * 4 pixels of adjacent and non-overlapping copies as trained vector; To the trained vector quadratic sum, sequence forms inceptive code book; Inceptive code book is practiced in the image construction trained vector training of choosing in image library; By adding up percentage frequency that each color occurs and the change color situation of adjacent pixel blocks, form color index histogram and main color transition matrix, using it as retrieval character; Utilize morphological images to process, highlight objective contour to extract topography interested zone; Utilize global color feature and local region of interest color characteristic Weighted Searching, obtain the region of interest area image.

2. method according to claim 1, is characterized in that, described calculating inceptive code book design comprises: the quadratic sum of each trained vector in the calculation training vector set, and arrange by ascending order; According to code book size N, the trained vector after sequence is divided into to the N section, select successively first code word of every section as initial code word, formation is of a size of the inceptive code book of N, adopt CL Algorithm for Training inceptive code book, obtain tone, saturation degree and brightness and form final code book, and output packet is containing the index value of the code word of the color look up table of N kind color.

3. method according to claim 1, is characterized in that, according to the percentage frequency that in the color look up table statistical picture, each codewords indexes occurs, forms the color index histogram; According to vector quantization codewords indexes table, image is carried out to Z-shaped scanning, the change color situation of statistics adjacent pixel blocks, form main color transition matrix.

4. method according to claim 1, is characterized in that, described topography interested extracted region specifically comprises: utilize opening and closing operation to carry out smoothing processing to image, retain the important profile of image and also remove noise; Setting threshold is removed tiny piecemeal, takes out initial row train value and the width of largest block and highly forms rectangular area; Utilize ranks number of pixels mould M to remove remainder, form the rectangle topography interested zone that the ranks pixel count is the integral multiple of M.

5. method according to claim 1, is characterized in that, utilizes the overall situation and local region of interest color characteristic to be weighted retrieval and be specially: the percentage frequency of the statistics overall situation and local each color, formation color index histogram H _a, H _b, H _a, H _b; The change color situation of statistics adjacent pixel blocks, form main color transition matrix D _a, D _b, D _a, D _b.Adopt Euclidean distance according to formula:

$\mathrm{Simi}1={\mathrm{\ω}}_1\sqrt{{(H_A-H_B)(H_A-H_B)}^T}+{\mathrm{\ω}}_2\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{A_{i,j}}-D_{B_{i,j}})}^2}$ Calculate overall similarity Simi1, wherein ω ₁for the histogrammic weight of global color index, ω ₂weight (ω for the main color transition matrix of the overall situation ₁, ω ₂∈ [0,1], and ω ₁+ ω ₂=1); According to formula:

$\mathrm{Simi}2={\mathrm{\ω}}_3\sqrt{{(H_a-H_b)(H_a-H_b)}^T}+{\mathrm{\ω}}_4\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(D_{a_{i,j}}-D_{b_{i,j}})}^2}$ Calculate local similarity Simi2, wherein ω ₃for the histogrammic weight of local color index, ω ₄weight (ω for the main color transition matrix in part ₃, ω ₄∈ [0,1] and ω ₃+ ω ₄=1); Synthetic similarity Similar=pSimi1+qSimi2, arrange Similar by ascending order, return to result for retrieval, p wherein, q ∈ [0,1] p+q=1.

Images