CN103440642B - Based on the strong and weak edge detection method of image of dot matrix neuron response space time information - Google Patents

Abstract

The present invention relates to a kind of strong and weak edge detection method of image based on dot matrix neuron response space time information.First the present invention utilizes the multi-direction Log-Gabor filter result of image in the visual field, the marginal information of reconstructed image; Then using reconstruction result as the neuronic input of dot matrix; Record the time that each neuron provides action potential first, formation time matrix; Then construct receptive field window to slide on time matrix, according to the variance after the sequential computed improved of each time element, and assignment is to window center point, thus obtains the variance matrix comprising neuron response time and spatial information; Afterwards aforementioned receptive field window is continued to slide on variance matrix, realize neuron side direction rejection characteristic spatially, obtain matrix of edge; Be finally result images by matrix of edge inverse mapping.Contemplated by the invention the space time information of dot matrix neuron response, can not only image border be detected, and effectively can reflect the strong or weak relation at edge.

Description

Based on the strong and weak edge detection method of image of dot matrix neuron response space time information

Technical field

The invention belongs to image processing field, relate to a kind of strong and weak edge detection method of image based on dot matrix neuron response space time information.

Background technology

The edge of image refers to the place of color or gray-scale value generation transition, it reflects the main information of image usually, therefore effective detection of the strong and weak marginal information of image, for follow-up image procossing and other relevant tasks, such as pattern-recognition, target following etc. are most important.Traditional Image Edge-Detection adopts roberts operator constant gradient method to process image, can detect the strong edge of image, but usually can lose weak details, sometimes can produce over-segmentation to image again.

Summary of the invention

For the problems referred to above, the present invention proposes a kind of strong and weak edge detection method of image based on dot matrix neuron response space time information, have recorded neuron corresponding to each pixel discharge time first in the method, utilize the order that in receptive field window, each neuron discharges first, and the marginal information of image is extracted in conjunction with the spatial character of neuron lateral inhibition, finally effectively can not only detect the edge of image, and can by the strong and weak Informational Expression at edge out.

The present invention is based on the strong and weak detection method in image border of dot matrix neuron response space time information, comprise the following steps:

Step (1) carries out 8 direction Log-Gabor filter process to image, and angle is respectively θ _i(θ _i=22.5 ⁰* i, i=0,1 ..., 7), then according to the marginal information of filter result reconstructed image.

Reconstruction result in step (1) is input to dot matrix neuron by step (2).

Step (3) records each neuronic discharge time first, then obtains corresponding time matrix.

The receptive field window that step (4) structure is one 3 × 3 slips over this time matrix, first the time element in this receptive field window is sorted, be weighted according to ranking results, calculate the improvement variance of discharge time first, and by its assignment to the central element of receptive field window.Successively identical process is carried out to each element in time matrix, then obtain variance matrix.

Step (5) makes the receptive field window of 3 × 3 of structure in step (4) slip over above-mentioned variance matrix, side direction is carried out to the neuron in receptive field window and suppresses process, need equally to carry out identical process successively to each element in variance matrix, then obtain matrix of edge.

Data inverse in matrix of edge is mapped to the scope of 0 ~ 255 by step (6), is shown as image afterwards, and this image is Image Edge-Detection result, and comprises strong and weak marginal information.

The feature that the present invention has is: the method considers the spatial information of temporal information that dot matrix neuron discharges first and receptive field window, the edge of image can not only be detected, and the strong or weak relation of image border can be provided, be embodied in following some:

1. in rim detection, introduce the time encoding mechanism of visual cortex Neural spike train pulse for visual stimulus, make use of dot matrix information neuronic discharge time first.

2. in rim detection, consider the existence of visual cortex neuron receptive field, utilize receptive field window to define the regional area of rim detection.

3. in rim detection, introduce the neuronic lateral inhibition mechanism of visual cortex, to each neuron establishment side in receptive field window to rejection characteristic.

4. the method for the present invention's employing and the working mechanism of vision system meet more, and acquired results can retain more detailed information compared with classic method, more can meet the subjective assessment of vision system.

5. the method acquired results that the present invention adopts can embody the strong and weak information of image border.

Accompanying drawing explanation

Fig. 1 is the inventive method process flow diagram.

Embodiment

Below in conjunction with accompanying drawing 1, the invention will be further described, and in the middle of Fig. 1, I_old (i, j) represents original input picture; f _k(i, j) (k=0,1 ..., 7) be θ through Log-Gabor wave filter with angle _i(θ _i=22.5 ⁰* i, i=0,1 ..., 7) filtered result; Neuron (i, j) represents the dot matrix neuron used; T (i, j) represents the time matrix recording discharge time first after dot matrix neuron models; D (i, j) represents the variance matrix through variance process; F (i, j) represents the matrix of edge after neuron side direction inhibiting effect; I_new (i, j) represents final result images.

The concrete steps of the inventive method are:

Step (1) make original image I_old (i, j) (i=1,2 ..., M; J=1,2 ..., N) carry out pre-service by Log-Gabor wave filter, acquisition angle is θ _i(θ _i=22.5 ⁰* i, i=0,1 ..., 7) 8 direction results, be designated as f _k(i, j) (k=0,1 ..., 7).Then the marginal information of formula (1) reconstructed image is utilized:

$\mathrm{result}(i,j)=\underset{k=0}{\overset{7}{\mathrm{\Σ}}}{f}_k(i,j)(i=\mathrm{1,2},...,M;j=\mathrm{1,2},...,N)---\left(1\right)$

Step (2) structure dot matrix neuron Neuron (i, j) (i=1,2 ..., M; J=1,2 ..., N), the reconstruction result result (i, j) in step (1), such as formula shown in (2), is input to dot matrix neuron Neuron (i, j) and processes by the neuronic model of each dot matrix.

$\left\{\begin{array}{c}\frac{\mathrm{dv}}{\mathrm{dt}}={0.04v}^2+5v+140-u+I_{\mathrm{ext}}\\ \frac{\mathrm{du}}{\mathrm{dt}}=a(\mathrm{bv}-u)\\ \mathrm{ifv}\≥v_{\mathrm{thresh}},\mathrm{thenv}\←u\←u+d\end{array}\right.---\left(2\right)$

Wherein, v is membrane potential of neurons, and u represents neuron membrane and recovers variable.I _extbe input stimulus, i.e. corresponding reconstruction result result (i, j), a, b, c and d are model parameter, v _threshfor threshold value, be set to a=0.02, b=0.2, c=-65, d=6, v _thresh=30.If v is more than or equal to v _thresh, neuron provides pulse, namely produces electric discharge, and v is reset to c simultaneously, and u is reset as u+d; Otherwise neuron does not provide pulse.

Step (3) record each dot matrix neuron Neuron (i, j) (i=1,2 ..., M; J=1,2 ..., N) discharge time first, thus obtain time matrix T (i, j) (i=1,2 ..., M; J=1,2 ..., N).

The receptive field window that step (4) structure is one 3 × 3 is corresponding with 3 × 3 regional areas of time matrix T (i, j), carries out ascending sort to time element corresponding in receptive field window, be designated as Templet (i), i=1,2 ... 9.Then to the variance after the time element computed improved of this receptive field window, shown in (3).

$\left\{\begin{array}{c}\mathrm{average}=\left(\underset{i=1}{\overset{9}{\mathrm{\Σ}}}\mathrm{templet}\left(i\right)\right)/9\\ \mathrm{sum}=\underset{i=1}{\overset{9}{\mathrm{\Σ}}}[{(\mathrm{templet}\left(i\right)-\mathrm{average})}^2*e^{(-i)}]\end{array}\right.---\left(3\right)$

Average is the mean value of each time element in receptive field window.Sum be discharge first sequence weighting after improvement variance, by its assignment to the central point of receptive field window.Receptive field window is slipped over time matrix T (i, j) successively, to each time element carry out same process just obtain variance matrix D (i, j) (i=1,2 ..., M; J=1,2 ..., N).

Step (5) utilize step (4) to construct 3 × 3 receptive field window, make itself and variance matrix D (i, j) 3 × 3 regional areas are corresponding, in receptive field window, realize neuronic side direction suppression mechanism according to formula (4).

$w_0=\left\{\begin{array}{ccc}{w}_0*e^{-(w_0/w_1)},& \mathrm{if}& w_0>w_1\\ w_0*e^{(w_1/w_0)},& \mathrm{if}& w_0>w_1\\ w_0,& \mathrm{if}& w_0=w_1\end{array}\right.$ (4)

$w_1=\left\{\begin{array}{ccc}{w}_1=w_1*e^{(w_0/w_1)},& \mathrm{if}& w_0<w_1\\ w_1=w_1*e^{-(w_1*w_0)},& \mathrm{if}& w_0>w_1\\ w_1=w_1,& \mathrm{if}& w_0=w_1\end{array}\right.$

W ₀the central element in receptive field window, w ₁represent the non-central element in receptive field window.Receptive field window is slipped over variance matrix D (i, j) successively, to each variance element carry out same process just obtain matrix of edge F (i, j) (i=1,2 ..., M; J=1,2 ..., N).

Matrix of edge data inverse is mapped to the scope of 0 ~ 255 by step (6) according to formula (5), and then is shown as image I_new (i, j), is the testing result comprising the strong and weak marginal information of image.

$I\_\mathrm{new}(i,j)=\frac{(F(i,j)-\min )}{(\max -\min )}*255---\left(5\right)$

Wherein min and max is respectively minimum value in matrix of edge F (i, j) and maximal value, and I_new (i, j) represents the testing result at the strong and weak edge of image.

In sum, contemplated by the invention following factor: some weak detailed information can be lost when (1) traditional edge detection method processes image, or carry out over-segmentation, thus original marginal information cannot be told; (2) traditional edge detection method adopts the method display image border of binaryzation usually, does not namely consider the strong or weak relation at edge; (3) traditional edge detection method adopts operator to carry out marginal information acquisition, and do not consider the physiological property of vision system, the result therefore drawn is difficult to be consistent with the subjective assessment of people usually.Therefore, the present invention is based on the space time information of dot matrix neuron response, the strong and weak detection method in the physiological image border of a kind of view-based access control model is proposed, make the image after processing can not only retain more detailed information, and the power of image border can be showed, result is consistent with the subjective assessment of people.

Claims (1)

1., based on the strong and weak edge detection method of image of dot matrix neuron response space time information, it is characterized in that the method comprises the steps:

Step (1) carries out 8 direction Log-Gabor filter process to image, and angle is respectively θ _i, θ _i=22.5 ⁰* i, i=0,1 ..., 7, then according to the marginal information of filter result reconstructed image;

Reconstruction result in step (1) is input to dot matrix neuron by step (2);

Described dot matrix neuron be Neuron (i, j) (i=1,2 ..., M; J=1,2 ..., N), model is as follows:

Wherein, v is membrane potential of neurons, and u represents neuron membrane and recovers variable, I _extbe input stimulus, i.e. corresponding reconstruction result result (i, j), a, b, c and d are model parameter, v _threshfor threshold value, be set to a=0.02, b=0.2, c=-65, d=6, v _thresh=30, if v is more than or equal to v _thresh, neuron provides pulse, namely produces electric discharge, and v is reset to c simultaneously, and u is reset as u+d; Otherwise neuron does not provide pulse; Step (3) records each neuronic discharge time first, then obtains corresponding time matrix;

The receptive field window that step (4) structure is one 3 × 3 slips over above-mentioned time matrix, first the time element in this receptive field window is sorted, be weighted according to ranking results, calculate the improvement variance of discharge time first, and by its assignment to the central element of receptive field window; Successively identical process is carried out to each element in time matrix, then obtain variance matrix;

The improvement variance of described discharge time first,

Average is the mean value of each time element in receptive field window, sum be discharge first sequence weighting after improvement variance, by its assignment to the central point of receptive field window, receptive field window is slipped over time matrix T (i, j) successively, same process is carried out to each time element and just obtains variance matrix D (i, j) (i=1,2 ..., M; J=1,2 ..., N); Step (5) makes the receptive field window of 3 × 3 of structure in step (4) slip over above-mentioned variance matrix, side direction is carried out to the neuron in receptive field window and suppresses process, need equally to carry out identical process successively to each element in variance matrix, then obtain matrix of edge; Described side direction suppresses to be treated to

W ₀the central element in receptive field window, w ₁represent the non-central element in receptive field window, receptive field window slipped over variance matrix D (i, j) successively, to each variance element carry out same process just obtain matrix of edge F (i, j) (i=1,2 ..., M; J=1,2 ..., N);

Data inverse in matrix of edge is mapped to the scope of 0 ~ 255 by step (6), is shown as image afterwards, and this image is Image Edge-Detection result, and comprises strong and weak marginal information.