CN104636753A - Region characteristic extraction method based on PCNN (Pulse Coupled Neural Network) neuron activation rate and cluster dispersion - Google Patents

Abstract

The invention discloses a region characteristic extraction method based on PCNN (Pulse Coupled Neural Network) neuron activation rate and cluster dispersion. The method comprises the following steps of 1, pretreating an original image and enabling the PCNN to correspond to the image; 2, partitioning a gray interval of 0-255 into N gradually-reduced gray intervals according to a target region; 3, obtaining neurons which are subjected to cluster activation in the current gray interval; 4, counting the quantity of the activated neurons in the region of neurons subjected to cluster activation, and judging the neurons which are subjected to lifting activation; 5, counting the quantity of the neurons which are subjected to lifting activation, and obtaining the cluster activation rate and the group dispersion; and 6, reading the next gray interval and repeating the steps 3-6 till the Nth gray interval. The method has the advantages of low computation complexity and good classification effect.

Description

A kind of Region Feature Extraction method of Based PC NN neuronal activation rate and group's dispersion

Technical field

The invention belongs to a kind of feature extracting method of sonar image target area, particularly relate to a kind of target area pixel value distribution characteristics of can extracting, the Region Feature Extraction method of Based PC NN neuronal activation rate and group's dispersion.

Background technology

Along with the development of science and technology, it is too far behind to catch up with the mankind to be that the mammal of representative remains machine for the recognition capability of image, therefore a large amount of researcher both at home and abroad constantly explores the visual system with simulated animal, although the research of neural network experienced by one and another low tide, the step of scientists never stops.Along with the proposition of third generation neural network PCNN, research for neural network enters a new chapter, Lindblad and Kinser in 2005 has set forth the artificial neural network based on mammalian visual cortical neuron in the second edition of its PCNN research treatise, people are a a progressive step in machine vision research, how have had again a lot of new achievement to the research aspect that image processes for animal brain.Such as Johnson proposes to utilize PCNN network two dimensional image to be converted to the thought of one dimension pulse train, and some researchers propose with PCNN image layered method.Document 1:Bo Yu; Liming Zhang.Pulse-Coupled Neural Networks for Contour and Motion Matchings.2004 IEEE TRANSACTIONS ON NEURAL NETWORKS, p1186-1201, the chain type conduction Activiation method of 2004, this section of document utilization PCNN carries out match cognization to the edge of Ideal graph.Document 2: Liu's Qing, Ma Yide. based on the PCNN images steganalysis new method of histogram vectors center of gravity. application of electronic technology, No.10,2006.Document 3: Liu's Qing, Xu Luping, Ma Yide, Zhang Hua. in conjunction with the PCNN Small object image new detecting method of gray level entropy conversion. Beijing Institute of Technology's journal, Vol.29, No.12,2009.When these two sections of document utilizations are composed, matrix carrys out Description Image feature for identifying.

Generally speaking, the paper utilizing PCNN to carry out image recognition can be divided into two large classes: a class builds more complicated PCNN structure, and the complicacy increasing network is innovated as one; The neuron number that Equations of The Second Kind, by statistics PCNN, group's activation occurs is analyzed, and carries out segmentation, the denoising of image with this.These ways are deposited not enough both ways, and one is have ignored PCNN neighborhood neuron to the dynamic effects of Master neuron, does not embody the main and auxiliary neuronic dynamic movement of neighborhood; Ignore PCNN many neurons there is previously activated neuron characteristic distributions simultaneously on the other hand, does not utilize the neuronic distribution character of activated in advance.

Summary of the invention

The object of this invention is to provide a kind of computation complexity little, good classification effect, the Region Feature Extraction method of Based PC NN neuronal activation rate and group's dispersion.

The present invention is achieved by the following technical solutions:

A Region Feature Extraction method for Based PC NN neuronal activation rate and group's dispersion, comprises following step:

Step one: gather original image, pre-service is carried out to original image, be partitioned into the profile of target area, neural network PCNN is corresponding with image, by corresponding with the pixel of image for central nervous unit, the neighborhood of central nervous unit is corresponding with neighborhood territory pixel point, the neuronic gray-scale value being input as pixel;

Step 2: by arranging the neuronic threshold parameter of PCNN, be divided between N number of gray area by the tonal range of 0 ~ 255 according to target area, arranges between gray area according to gray-scale value order from big to small;

Step 3: read between a current kth gray area, obtains occurring in current gray level interval the neuron that group activates, and pixel value corresponding to neuron that group activates occurs in the interval corresponding tonal range of current gray level;

Step 4: statistics group occurs and activates excited target neuron number λ in neuronic field, recalculate and the value that group activates the neuronic internal activity item of excited target in neuronic field occurs, if the value of the neuronic internal activity item of excited target is positioned at the interval corresponding tonal range of current gray level, then current excited target neuron is activated in advance neuron, records the coordinate of activated in advance neuron corresponding pixel points;

Step 5: statistics activated in advance neuron number λ ', obtains group activity ratio and group's dispersion:

Wherein x, y are the coordinates of activated in advance neuron corresponding pixel points between each gray area,

Step 6: make the pixel point value that the generation group in current gray level interval activates and activated in advance neuron is corresponding be 0, make k=k+1, repeats step 3 to step 6, until k=N.

Beneficial effect:

The refinement of the present invention neuron exciting phenomena of PCNN, success is activated generation group and the previously activated neuron of excited target is separated, the neuronic activity ratio of activated in advance is utilized to characterize the grey value profile characteristic of target area on the one hand, add up pixel coordinate corresponding to activated in advance neuron between each gray area on the other hand, the pixel in target area with similar activation characteristic is connected analysis.

The present invention utilizes the neuronic distributed architecture of PCNN neighborhood, flexibly target area is divided into pixel region among a small circle, characterizes the pixel value distribution characteristics of these pixel regions among a small circle with previously activated neuron; The region with similar activating property is considered jointly, take into account the pixel value gamma characteristic of neighborhood gradient characteristics and more large regions scope, PCNN activated in advance neuron distribution discreteness in a wider context can be added up, therefore describe the feature of target area at microcosmic and macroscopic view two kinds of different angles.Compared to other texture characteristic extracting method, when utilizing context of methods extraction frogman and fish target signature for classifying, computation complexity is less, classifying quality is better.

Accompanying drawing explanation

Fig. 1 is the corresponding method of PCNN neuron and sonar image;

Fig. 2 is the gray level image of target region;

Fig. 3 is the target image be partitioned into;

Fig. 4 is the pixel activated first time, and respective pixel is worth between maximum gray area;

Fig. 5 is the pixel region affected by activation neuron;

Fig. 6 is the pixel that activated in advance phenomenon occurs in affected pixel region;

Fig. 7 is the suppressed region no longer occurring to activate afterwards;

Fig. 8 is that un-activation neuron Q affects by having activated neuron P in neighborhood, and previously activated schematic diagram occurs;

Fig. 9 activates neuron statistical graph for group;

Figure 10 is group's activity ratio statistical graph;

Figure 11 is group's dispersion statistical graph;

Figure 12 is the sonar image of fish target;

Figure 13 is the sonar image of frogman's target;

Figure 14 is the bulk treatment process flow diagram of the inventive method.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further details.

The present invention is achieved in that

1. pair original image carries out pre-service, splits, calibrates target area, and PCNN is corresponding with sonar image, and central nervous unit is corresponding with the pixel of image, the gray-scale value of neuronic input and pixel, and the neighborhood of central nervous unit is corresponding with neighborhood territory pixel point;

2. PCNN neuron threshold value correlation parameter is set, is divided into N number of interval by autotelic for the gray level of 0 ~ 255;

3. the grey level range divided in statistic procedure 2, according to from the relatively large interval of gray-scale value to the interval statistics (between each statistics gray area) that gray-scale value is relatively little, judges whether the neuron in current gray level interval group occurs and activates phenomenon;

4., according to the neuronic character of PCNN, activated neuron and un-activation neuron in neighborhood can be encouraged to activate.Therefore influenced neuronic U value will be recalculated;

5. judge whether affected un-activation neuron again meets activation condition and activate (this is a kind of activated in advance phenomenon).If met, the neuron generation activated in advance that this pixel is corresponding, records coordinate and the quantity of previously activated neuron corresponding pixel points;

6. make the pixel point value that all generation groups activate and the previously activated neuron of excited target is corresponding be 0, no longer activate;

7. get back to step 3, add up the group's activating property between lower picture gray area, according to the previously activated character of order statistics neuron from step 3 to step 6, until all intervals of gray-scale value in 0 ~ 255 scope all add up complete;

8. add up neuron number and previously activated neuron number that group's activation occurs between each gray area, obtain group activity ratio.Add up the pixel coordinate that activated in advance neuron between each gray area is corresponding, obtain group dispersion.Using group's activity ratio and group's dispersion as the eigenwert describing provincial characteristics.

Implementation process of the present invention is described below, mainly in conjunction with Figure 14, the present invention is described in more detail:

The fundamental formular of PCNN is as follows:

$F_{\mathrm{ij}}\left[n\right]=F_{\mathrm{ij}}[n-1]\exp (-t{\mathrm{\α}}_F)+V_F\underset{\mathrm{kl}}{\mathrm{\Σ}}{M}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}[n-1]---\left(1\right)$

$L_{\mathrm{ij}}\left[n\right]=L_{\mathrm{ij}}[n-1]\exp (-t{\mathrm{\α}}_L)+V_L\underset{\mathrm{kl}}{\mathrm{\Σ}}{M}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}[n-1]---\left(2\right)$

U _ij[n]＝F _ij[n](1+βL _ij[n]) (3)

T _ij[n]＝exp(-tα _T)T _ij[n-1]+V _TY _ij[n-1] (4)

$Y_{\mathrm{ij}}\left[n\right]=\mathrm{step}\left(U_{\mathrm{ij}}\right[n]-T_{\mathrm{ij}}[n-1\left]\right)=\left\{\begin{array}{c}1,U_{\mathrm{ij}}\left[n\right]\&GreaterEqual;T_{\mathrm{ij}}\left[n\right]\\ 0,U_{\mathrm{ij}}\left[n\right]<T_{\mathrm{ij}}\left[n\right]\end{array}\right.---\left(5\right)$

In above formula:

F _ij[n] represents neuron primary input.

S _ijrepresent the image pixel gray level value that primary input central point is corresponding.

L _ij[n] represents auxiliary input, and its value represents the auxiliary input of primary input neighborhood.

U _ij[n] is internal activity item.

T _ij[n] is dynamic threshold.

M _ijkland W _ijklbe neuronic link weight matrix, represent neuronic and experience the visual field.

Y _kl[n-1] is the pulse Output rusults of pulse-coupled neural networks, and value is 1 (activation) and 0 (un-activation).

Ij represents the transverse and longitudinal coordinate of primary input region inconocenter vegetarian refreshments.Kl represents the transverse and longitudinal coordinate of auxiliary neighborhood territory pixel point.

V _f, V _land V _tbe the affecting parameters of neighborhood, β is the interaction strength between primary input F road and auxiliary input L road.

N=1,2 ..., n is the multiple in minimum sampling period, represents which time dynamic change current formula is in, hereinafter referred to as iterations.

Below in conjunction with accompanying drawing, implementation process of the present invention is described in detail:

1. as shown in Figure 1, PCNN is corresponding with sonar image, namely central nervous unit is corresponding with the pixel of image, and neighborhood neuron is corresponding with neighborhood territory pixel point, the gray-scale value of neuron input and pixel.Carry out pre-service to original image, fundamental purpose is smoothing denoising, is partitioned into target area, obtains Fig. 2, and PCNN corresponds to the image after process; Here can split objective contour, by the gray-scale value zero setting of background, eliminate the impact of ground unrest, obtain Fig. 3.

2. by PCNN fundamental formular abbreviation:

F when iteration starts _ij[n] equals PCNN central nervous unit corresponding grey scale value, i.e. F _ij[n-1]=S _ij, all neurons are all in holddown, i.e. Y _ij[n-1]=0, now the activation characteristic of central nervous unit is by S _ijwith attenuation parameter α _fimpact, if the constant removed in formula (1) is biased S _ij, formula (1) is deformed into:

F _ij ^*[n]＝S _ijexp(-tα _F) (6)

When judging whether un-activation neuron occurs previously activated first, the neuron that main consideration occurs to activate is on the impact of central nervous unit, and therefore formula (2) abbreviation is:

$L_{\mathrm{ij}}\left[n\right]=V_L\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{W}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}[n-1]---\left(7\right)$

The impact of PCNN neighborhood neuron on central nervous unit is a dynamic process, after the initial activation of neighborhood neuron, and Y _ijits peripheral nerve unit of appreciable impact can there is activated in advance in the change of [n-1].Therefore formula (3) is deformed into two parts by the present invention:

Part I: U _ij ^*[n]=F _ij ^*[n] (8)

Part II: U _ij[n]=U _ij ^*[n]+β F _ij ^*[n] L _ij[n] (9)

Judge whether neuron group occurs when activating, and all neurons are all in holddown, i.e. Y first _ij[n-1]=0.Therefore formula (4) abbreviation is:

T _ij[n]＝exp(-tα _T)T ₀(10)

Wherein T ₀for initial threshold.

PCNN fundamental formular (9) and formula (10) correlation parameter are set, close gray-scale value can be activated according to formula (5) in identical sampling instant simultaneously; The gray level of 0 ~ 255 be divided into several interval, between each gray area, comprise the neuron that can simultaneously activate;

3., according between the gray area divided in step 2, according to the order that gray-scale value corresponding between gray area is descending, between each Iterative statistical gray area, judge according to formula (5) the PCNN neuronal activation situation that sonar image pixel is corresponding.Figure 4 shows that first time iteration, namely the pixel that the neuron of group's activation is corresponding occurs in maximum gray area;

4., according to the neuronic exciting phenomena of PCNN, activated neuron and un-activation neuron in neighborhood can be encouraged to activate, as shown in Figure 8.Therefore formula (9) is utilized to recalculate influenced neuronic U _ij[n] value, Figure 5 shows that the pixel that excited target neuron is corresponding;

5. judge whether affected un-activation neuron again meets activation condition and activate (this is a kind of activated in advance phenomenon), if met, the neuron generation activated in advance that this pixel is corresponding, record coordinate and the quantity of previously activated neuron corresponding pixel points, Figure 6 shows that the pixel that activated in advance neuron is corresponding;

6. make the pixel point value that all generation groups activate and the previously activated neuron of excited target is corresponding be 0, no longer activate, as shown in Figure 7;

7. return step 3 and enter next iteration, neuron, excited target neuron and the neuronic situation of activated in advance is activated according to the group between the lower gray area of the order statistics from step 3 to step 6, until all intervals of gray-scale value in 0 ~ 255 scope all add up complete, obtain the time pulse sequence figure shown in Fig. 9, in figure, horizontal ordinate is iterations, ordinate be between gray area corresponding to each iteration in there is the neuron number that group activates;

8. by successfully there is previously activated number λ ' in the statistics neuron number λ of excited target and these neurons in the present invention, obtains activating neuron to the neuronic group's activity ratio of neighborhood (hereinafter referred to as " group's activity ratio "), and it is defined as follows:

Add up group's activity ratio that activated in advance neuron between each gray area is corresponding, can obtain group's activity ratio vector as shown in Figure 10, in figure, horizontal ordinate is iterations, the group activity ratio of ordinate corresponding to each iteration between gray area.

In addition, the present invention proposes the concept (being called for short " group's dispersion " below) that a kind of PCNN " activates group's dispersion ", is used for describing previously activated neuron distribution dispersion spatially.The reason proposing it is: group's activity ratio characterizes the neuronic fillip of neighborhood and texture properties, but previously activated neuron is not embodied at wider spatial distribution characteristic, therefore the concept proposing " group's dispersion " characterizes this feature, and formula (12) is the definition of group dispersion:

Wherein x, y are the coordinates of activated in advance pixel between each gray area.

$x_0=\frac{1}{{\mathrm{\&lambda;}}^{\&prime;}}\underset{i=1}{\overset{{\mathrm{\&lambda;}}^{\&prime;}}{\mathrm{\&Sigma;}}}{x}_i,y_0=\frac{1}{{\mathrm{\&lambda;}}^{\&prime;}}\underset{i=1}{\overset{{\mathrm{\&lambda;}}^{\&prime;}}{\mathrm{\&Sigma;}}}{y}_i---\left(13\right)$

Add up group's dispersion that activated in advance neuron between each gray area is corresponding, obtain the group's dispersion vector shown in Figure 11, in figure, horizontal ordinate is iterations, the group dispersion of ordinate corresponding to each iteration between gray area.The concept of group's dispersion compensate for the neighborhood limitation of activity ratio sign, embodies and activates the spatial distribution characteristic of neuron in view picture target zone.

Application context of methods, can classify the sonar image (and other image) shown in Figure 12 and Figure 13.Method is: carried out mating (matching process of group's dispersion vector is identical) by group's activity ratio vector of to be sorted two targets, matching degree is defined as follows:

$\mathrm{\&eta;}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\frac{1}{x_{1i}-x_{2i}}$

X in formula _1iand x _2irepresent the value of the above-mentioned proper vector correspondence position of two targets respectively, N represents gray scale interval number.

Prove more than a matching degree order of magnitude higher than the matching degree of inhomogeneity target of the similar target of the present invention through test, there is good classifying quality.

Claims (1)

1. a Region Feature Extraction method for Based PC NN neuronal activation rate and group's dispersion, is characterized in that, comprises following step:

Step one: gather original image, pre-service is carried out to original image, be partitioned into the profile of target area, neural network PCNN is corresponding with image, by corresponding with the pixel of image for central nervous unit, the neighborhood of central nervous unit is corresponding with neighborhood territory pixel point, the neuronic gray-scale value being input as pixel;

Step 2: by arranging the neuronic threshold parameter of PCNN, be divided between N number of gray area by the tonal range of 0 ~ 255 according to target area, arranges between gray area according to gray-scale value order from big to small;

Step 3: read between a current kth gray area, obtains occurring in current gray level interval the neuron that group activates, and pixel value corresponding to neuron that group activates occurs in the interval corresponding tonal range of current gray level;

Step 4: statistics group occurs and activates excited target neuron number λ in neuronic field, recalculate and the value that group activates the neuronic internal activity item of excited target in neuronic field occurs, if the value of the neuronic internal activity item of excited target is positioned at the interval corresponding tonal range of current gray level, then current excited target neuron is activated in advance neuron, records the coordinate of activated in advance neuron corresponding pixel points;

Step 5: statistics activated in advance neuron number λ ', obtains group activity ratio and group's dispersion:

Wherein x, y are the coordinates of activated in advance neuron corresponding pixel points between each gray area,

Step 6: make the pixel point value that the generation group in current gray level interval activates and activated in advance neuron is corresponding be 0, make k=k+1, repeats step 3 to step 6, until k=N.