CN104933722A - Image edge detection method based on Spiking-convolution network model - Google Patents

Abstract

The invention discloses an image edge detection method based on a Spiking-convolution network model, which belongs to the technical field of image processing and solves the problem that a method in the prior art only simulates the spatial hierarchical structure of a biological nervous system, but lacks interpretation to time characteristics. The image edge detection method disclosed by the invention comprises the following steps: creating the Spiking-convolution network model of a convolution structure having an input layer, a Spiking-convolution layer and an output layer based on an information processing connection manner of a visual hierarchical structure; using a Laplace-Gauss operator and a Gauss difference operator as a filter of the Spiking-convolution layer of the created Spiking-convolution network model of the convolution structure to form a Spiking-convolution algorithm based on the operators; obtaining an image, encoding gray value pixels of the image into Spiking neurons to serve as the input layer of the Spiking-convolution network model; applying the Spiking-convolution algorithm based on the operators to the Spiking-convolution network model, carrying out pulse convolution on the input layer, and then recreating and outputting the edge of the image according to a Spiking threshold ignition model. The image edge detection method disclosed by the invention is applied to image pre-processing, characteristic extraction and edge detection and relates to neural networks, machine learning and Deep Learning.

Description

A kind of method for detecting image edge based on Spiking-convolutional network model

Technical field

Based on Spiking ?the method for detecting image edge of convolutional network model, be applied to Image semantic classification, feature extraction, rim detection, relate to neural network, machine learning, Deep Learning, belongs to the technical fields such as image procossing.

Background technology

In the image processing problem of reality, the outline map of image is as a kind of essential characteristic of image, in the image procossing being frequently applied the feature interpretation of higher level, image recognition, Iamge Segmentation, image enhaucament and compression of images etc. and analytical technology, thus can be for further analysis and understand to image.It is in image recognition, Iamge Segmentation, has and apply comparatively widely in the field of image enhaucament and compression of images etc., is also their basis.

Image border is one of the most basic feature of image, often carries the most information of piece image.And marginal existence is in the irregular structure and not steady phenomenon of image, also singular points or catastrophe point place is namely present in, these points give the position of image outline, these profiles are usually then our more very important characteristic conditions required when image procossing, and this just needs us to detect and extract its edge image to piece image.Edge detection algorithm is then one of classical technical barrier in image processing problem, and its solution carries out high-level feature interpretation, identification and understanding etc. for us great impact; Again because rim detection has very important practical value in many aspects, so people are being devoted to research and are solving how to construct the problem with good nature and effective edge detection operator always.Under normal conditions, the singular point in signal or catastrophe point can be thought the marginal point in image by us, and near it, the situation of change of gray scale can reflect from the gradient of its adjacent image point intensity profile.

The essence of rim detection adopts certain algorithm to extract the boundary line between objects in images and background.Edge definition is that in image, zone boundary jumpy occurs gray scale by we.The situation of change of gradation of image can reflect by the gradient of gradation of image distribution, and therefore we can obtain edge detection operator with topography's differential technology.Classical edge detection method is to construct edge detection operator to certain small neighbourhood of pixel in original image.

Being described as the Spiking neural network of " third generation neural network model ", is effectively to simulate the power system that information between biological neuron transmits in time continuously.This model adopts time encoding mode organizational information, can encoding mechanism in Reality simulation biology, the precise time adopting pulse to provide is encoded, no matter provide frequency to coding mode information closer to actual biological nervous system than the pulse of traditional neural network, be that processing power and computing velocity are obtained for tremendous increase.Research shows, Spiking neuron inherently possesses the Nonlinear Processing ability to external input information, and compare front two generations's artificial neural networks, Spiking possesses stronger computing power.Spiking neuron models have more research in the subject such as biological, neural, and are still in the starting stage in the application of engineering field.

Deep Learning is the new study hotspot of in machine learning field, effectively can carry out decryption treatment scheme by mimic biology brain spatial hierarchy.Because it adapts to the background of large data age, and be successfully applied to the field such as semanteme, image, make the convolutional neural networks deposited for many years again start smart field new round research tide.As the Disciplinary Frontiers of machine learning, Deep Learning simulates the abstract cognition of the continuous iteration of biological nervous system and learning process, provides more reasonably Neurobiology foundation for optimizing neural computation model.But because convolutional neural networks only simulates the spatial hierarchy of biological nervous system, lack the explanation to temporal characteristics, remain and adopt the mode of discretize to process, so not yet fundamentally solve timing analysis problems, still there is huge room for improvement.

Neural network is the research topic of one kind of multiple subject crossing, along with deepening continuously and the continuous expansion of range of application of application technology, it is used to solve the difficult problem that a lot of Traditional Scientific can't resolve, for people learn the world, find unknown category, strengthen modern science scientific research level thus bring active influence with science and technology drive yield-power, increase people economic development amplitude, become the internationally recognized most advanced and sophisticated advanced subject in scientific domain direction.To the exploration of Spiking neural network, be the important step promoting nerual network technique application, there is high practical engineering application and be worth.

Summary of the invention

The present invention is directed to the deficiencies in the prior art part and provide a kind of method for detecting image edge based on Spiking-convolutional network model, the spatial hierarchy of biological nervous system and the explanation feature of temporal characteristics can be simulated, Spiking biomechanism is adopted to process, effectively can catch space time information, have more bio-imitability and space-time characterisation than prior art, thus the edge of image can be applied in the feature interpretation of higher level.

To achieve these goals, the technical solution used in the present invention is:

Based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, following steps:

(1) the information processing connected mode of view-based access control model hierarchy, be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?convolutional network model;

(2) the Spiking-convolutional network model of convolutional coding structure will built, use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?convolution algorithm;

(3) obtaining image, is Spiking neuron by image intensity value pixel coder, as Spiking ?the input layer of convolutional network model;

(4) the Spiking ?convolution algorithm based on operator is applied to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the edge of reconstruct output image.

Further, in described step (1), be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?the concrete steps of convolutional network model as follows:

(11) set up one " Shu Ru Ceng ?Spiking ?Juan Ji Ceng ?output layer " pattern 3 layers of Spiking ?convolutional neural networks structure;

(12) according to Spiking ?the data characteristics of convolutional network structure and two dimensional image, set the two-dimensional matrix of and pretreatment image identical dimensional, and make each pixel be mapped to one by one Spiking ?input layer in convolutional neural networks structure;

(13) Spiking ?in convolutional network structure, the vision system that simulation is biological, the function of emulation receptive field, by Spiking ?input layer in convolutional neural networks structure to Spiking ?Spiking in convolutional neural networks structure ?the connected mode of convolutional layer be reduced to subregion and connect, obtain the convolution partially connected mode of Spiking nerve impulse;

(14) after obtaining the convolution partially connected mode of Spiking nerve impulse, again to Spiking ?in convolutional neural networks structure Spiking ?each receptive field of convolutional layer adopt weights to share, the weights that all receptive fields adopt are all identical, namely the wave filter of each Spiking-convolutional layer can repeatedly act in each region, the result of input signal being carried out to convolution constitutes input signal feature, thus extract the local feature of signal, the wave filter of each Spiking-convolutional layer is identical, namely identical parameter is shared, comprise identical weight matrix and bias term, final generation Spiking ?convolutional network model.

Further, in described step (2), use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?the concrete steps of convolution algorithm as follows:

(21) according to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Laplce's Gaussian filter;

(22) according to difference of Gaussian (DOG) function, difference of Gaussian (DOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Difference of Gaussian filter;

(23) by the gradient template in step (21) and step (22), as the convolution kernel in Spiking-convolutional network model, with identical template repeat function in each receptive field region, form the Spiking-convolution algorithm based on operator.

Further, in described step (21), according to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and the concrete steps with the equal effect of Laplce's Gaussian filter are:

(211) first with Gaussian function to the smoothing filtering of image;

(212) smoothing filtered image is carried out Laplace computing;

(213) after will carrying out Laplace computing, null point is as marginal point.

Further, in described step (3), be Spiking neuron by image intensity value pixel coder, as Spiking ?the concrete steps of input layer of convolutional network model as follows:

(31) by the pixel of input picture, image gray processing process is carried out, in pixel value range is limited to [0,255] interval;

(32) Spiking time delay encoding operation is carried out to each gray-scale value of image, adopt the time window of T=[0,255], adopt formula t _j=T-c × x _j, T represents time window, and c represents time constant, x _jrepresent input original image gray-scale value, t _jthe duration of ignition after presentation code, obtain the concrete Spiking pulse firing time by gray-scale value, when input picture gray scale stimulates larger, so its corresponding Spiking neuron firing intensity is stronger, and the pulse firing time of performance more early;

(33) input layer in Spiking-convolutional network model selects the two-dimensional matrix with image dimension formed objects, each expression Spiking neuron in matrix, forms input Spiking time of its input layer according to encode each Spiking neuron firing time of obtaining of the time delay of step (32).

Further, in described step (4) by based on operator Spiking ?convolution algorithm apply to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the concrete steps at the edge of reconstruct output image are as follows:

(41) the Spiking-convolution algorithm based on operator is acted on the Spiking neuron of input layer, namely in input layer Spiking neuron, using the convolution kernel of DOG/LOG wave filter as Spiking-convolutional network, carry out Spiking-convolution, its concrete Convolution Formula is: $X_j^k=f(\underset{i\∈M_j}{\Σ}T\left(X_i^{k-1}\right)*{\mathrm{Kernel}}_{ij}^k+B^k),$ the value of Spiking-convolutional layer, be the value of input layer, T is Spiking neuron coded system, and k represents which layer of network, and Kernel is convolution kernel, and i, j represent the corresponding network number of plies, and each characteristic pattern can be produced by different convolution kernels, and M _jfor a selection of input feature vector figure, every one deck has unique skew B, and f activation function is Spiking pulse voltage function ${\mathrm{\ϵ}}_{ij}\left(t\right)=\frac{t-t_i^f{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}}\exp (1-\frac{t-t_i^f{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}})H(x-t_i^f{-\mathrm{\Δ}}_{ij}^{ax}),$ ε _ijt (), as f activation function, wherein τ is time constant, axonal delay, step function, be the input layer duration of ignition, t is time shaft, and α x is aixs cylinder constant;

(42), after Spiking-convolution, the neuronic voltage curve of each post-synapse Spiking is produced immediately;

(43) according to the threshold value fire model SRM of Spiking, setting threshold value, the neuronic voltage curve of record post-synapse Spiking arrives the precise time point of threshold value, thus create the neuronic igniting sequential of Spiking post-synapse, wherein there are kind of special circumstances, namely be Spiking neuron in whole time window, when its maximum magnitude of voltage does not arrive threshold value, recording its Spiking neuronic duration of ignition is 0;

(44) the igniting sequential of Spiking post-synapse neuron generation, as the numerical point of the Spiking-convolutional layer in Spiking-convolutional neural networks;

(45) by step (44) by decoding and reconstituting, Spiking post-synapse neuronic duration of ignition directly as image intensity value, and is normalized image;

(46) after being normalized by image, carry out binaryzation operation to output layer image, producing output layer, is namely the marginal information of image.

Compared with prior art, the invention has the advantages that:

One, convolution is carried out to Spiking pulse, optimize internetwork connection mode, reduce weight complexity, improve computing velocity;

Two, adopt convolution biomechanism, simulating human vision system message processing flow, carries out abstract layered characteristic extraction;

Three, there is the processing mode of Spiking sequential mechanism, can efficient capture temporal information.Have the convolutional network framework of Space Rotating shift invariant, highly bionical thing brain system function, catches space time information simultaneously;

Four, by build the less 3 layers of Spiking of parameter ?convolutional neural networks framework, strengthen the arithmetic capability of computation model;

Five, by input picture being converted into neuron firing burst length sequence and Spiking ?convolutional network Structure Calculation mechanism, the processing horizontal of significant increase nonlinear data;

Six, by the template gradient based on operator, carry out Spiking ?convolution, by Laplce's Gaussian function (LOG) and difference of Gaussian function (DOG), can to the smoothing filtering of view data, simultaneously according to the singular point in signal and catastrophe point as the marginal point in image, reflect the situation of change of gray scale near it from the gradient of its neighbor intensity profile;

Seven, adopt weights to share, wave filter repeat function is in each receptive field, and under Spiking mechanism, operate time pulse voltage kernel function, carries out convolution operation to signal, drastically increase arithmetic speed;

Eight, adopt Spiking threshold value fire model to carry out modeling, greatly simulate biological neuron mechanism, its most essential data characteristics can be caught to signal pulse;

Nine, Spiking ?convolutional encoding mechanism, by the view data of complexity, rarefaction representation is two-dimensional network input layer, can process picture signal complicated and changeable in real life;

Ten, decoding and reconstituting, is converted into last image and exports, restore the edge feature of image essence by final neuron pulse firing sequential.

11, the present invention has organically combined Spiking neuromechanism, specifically propose a kind of have Spiking that good spatial characterization characteristic and temporal information transmit ?convolutional neural networks model, relate to the convolution process of Spiking time pulse, partially connected, weights are shared, reduce the Connecting quantity of network, optimized network, improves computing power.

Accompanying drawing explanation

Fig. 1 is method flow schematic diagram of the present invention;

Fig. 2 is overall flow schematic diagram of the present invention;

Fig. 3 be Spiking of the present invention ?convolutional network structural design schematic diagram;

Fig. 4 is original image of the present invention coding Spiking neuron process flow diagram;

Fig. 5 is partially connected schematic diagram between cynapse of the present invention;

Fig. 6 is that weights of the present invention share principle schematic;

Fig. 7 be network Spiking of the present invention ?the network connection diagram of convolutional layer;

Fig. 8 be the present invention is based on operator Spiking ?convolution method process flow diagram;

Fig. 9 is Spiking time delay of the present invention coding schematic diagram;

Figure 10 is Spiking threshold value igniting model schematic in the present invention;

Figure 11 is the Spiking-convolution mechanism flow through a network figure that the present invention's entirety uses.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described.

Based on a method for detecting image edge for Spiking-convolutional network model, first based on the information processing connected mode of human brain visual hierarchy, build the Spiking network model of a convolutional coding structure.First in conjunction with sequential Spiking mechanism and the convolution pattern of spatial abstraction ability.Use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer again, form a Spiking-convolution method based on operator.Be finally Spiking neuron by picture signal sparse coding, input Spiking-convolutional network structure, carries out the mode treatment of three-decker, noise-removed filtering, decoding and reconstituting, produces image edge information.

First need build Spiking ?convolutional network model, comprising: the definition of Spiking network input/output relation, convolutional network hierarchical structure, neural network on-link mode (OLM), data homogenization and coded system.Concrete steps are as follows:

(1) the information processing connected mode of view-based access control model hierarchy, be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?convolutional network model, namely the Spiking neuron with time sequence information is utilized, substitute traditional neural unit, the spatial extraction structure of construction convolutional network, use its time treatment mechanism, carry out pulse convolution, be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?the concrete steps of convolutional network model as follows:

(11) biological convolution mechanism is adopted, improve Spiking network structure, comprise local to connect, weights are shared, and sparse coding, to the gradation of image Value Data of catching, adopt time delay encoding scheme, be encoded to Spiking nerve impulse firing information, set up one " Shu Ru Ceng ?Spiking ?Juan Ji Ceng ?output layer " pattern 3 layers of Spiking ?convolutional neural networks structure, neuron wherein all adopts Spiking neuronal mechanism;

(12) according to Spiking ?the data characteristics of convolutional network structure and two dimensional image, set the two-dimensional matrix of and pretreatment image identical dimensional, and make each pixel be mapped to one by one Spiking ?input layer in convolutional neural networks structure;

(13) Spiking ?in convolutional network structure, the vision system that simulation is biological, the function of emulation receptive field, be reduced to subregion by the input layer in Spiking ?convolutional neural networks structure to the connected mode of the Spiking ?convolutional layer in Spiking ?convolutional neural networks structure to connect (in linear one-dimentional structure, show as and reduce number of connection), because interlayer local space correlativity, a neurode upper strata neurode close with it of adjacent every layer is connected, namely local connects, (biological vision system is the abstract treatment scheme of a highl stratification to obtain the convolution partially connected mode of Spiking nerve impulse, for the receptive field function of visual cortex, Spiking ?convolutional neural networks can simulate the sparse method for expressing of this subregion, correlated characteristic is extracted to specific region),

(14) after obtaining the convolution partially connected mode of Spiking nerve impulse, again to Spiking ?in convolutional neural networks structure Spiking ?each receptive field of convolutional layer adopt weights to share, the weights that all receptive fields adopt are all identical, namely the wave filter of each Spiking-convolutional layer can repeatedly act in each region, the result of input signal being carried out to convolution constitutes input signal feature, thus extract the local feature of signal, the wave filter of each Spiking-convolutional layer is identical, namely identical parameter is shared, comprise identical weight matrix and bias term, final generation Spiking ?convolutional network model.

(2) the Spiking-convolutional network model of convolutional coding structure will built, use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?convolution algorithm; Use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?the concrete steps of convolution algorithm as follows:

(21) according to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Laplce's Gaussian filter; According to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and the concrete steps with the equal effect of Laplce's Gaussian filter are:

(211) first with Gaussian function to the smoothing filtering of image;

(212) smoothing filtered image is carried out Laplace computing;

(213) after will carrying out Laplace computing, null point is as marginal point.

If f (x, y) is original image, G (x, y) is Gaussian function, and wherein Gaussian function prototype is: $G(x,y)=G(x,y,\mathrm{\δ})=\frac{1}{2{\mathrm{\π\δ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\δ}}^2}),$ Wherein δ is variance, and x, y are the two-dimensional coordinate value that in image, each pixel is corresponding.Carry out Laplace computing again after adopting this operator smoothed image, be namely LOG operator, be designated as as an operator, formula ${\▿}^2\[G(x,y,\mathrm{\δ})\]=\frac{1}{{\mathrm{\π\δ}}^4}\left(\frac{x^2+y^2}{2{\mathrm{\δ}}^2}\right)\exp \[-\frac{x^2+y^2}{2{\mathrm{\δ}}^2}\]={\▿}^{2}G=LOG(x,y,\mathrm{\δ}),$ Wherein δ is variance, the process adopting Gaussian function to carry out Spiking-convolution is exactly to Image Low-passed filtering, then employing Laplace operator is equal to the process to image high-pass filtering, integrating LOG operator is exactly bandpass filtering, in method normal adopt the LOG gradient template of two different δ of 5*5 be:

$\left(\begin{array}{ccccc}0& 0& -1& 0& 0\\ 0& -1& -2& -1& 0\\ -1& -2& 16& -2& -1\\ 0& -1& -2& -1& 0\\ 0& 0& -1& 0& 0\end{array}\right)\left(\begin{array}{ccccc}-2& -4& -4& -4& -2\\ -4& 0& 8& 0& -4\\ -4& 8& 24& 8& -4\\ -4& 0& 8& 0& -4\\ -2& -4& -4& -4& -2\end{array}\right).$

(22) according to difference of Gaussian (DOG) function, difference of Gaussian (DOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Difference of Gaussian filter; Difference of Gaussian (DOG) function is equivalent to the bandpass filter that can be removed the every other frequency information except those are retained the frequency of getting off in original image, DOG is the wavelet mother function of an empty total value, it deducts a wide Gauss from a narrow Gauss, is that of mexican hat wavelet (i.e. LOG) is similar to.The gradient modules in step (21) has been used equally in the present invention, and the Gaussian matrix difference value of two different δ, be set as δ respectively ₁=0.7 and δ ₂when=1.7, effect is best.

(23) by the gradient template in step (21) and step (22), as Spiking ?convolution kernel in convolutional network model, with identical template repeat function in each receptive field region, as Spiking ?the convolution Kernel function of convolutional layer, formed based on operator Spiking ?convolution algorithm.

(3) image is obtained, the image obtained is identical with the two-dimensional matrix set in step (12), be Spiking neuron by image intensity value pixel coder, be Spiking neuron using image intensity value pixel coder as Spiking ?the input layer of convolutional network model; Be Spiking neuron using image intensity value pixel coder as Spiking ?the concrete steps of input layer of convolutional network model as follows:

(31) input layer of Spiking-convolutional network model is a two-dimentional Spiking neuron matrix, by the pixel of input picture, carries out image gray processing process, in pixel value range is limited to [0,255] interval;

(32) Spiking time delay encoding operation is carried out to each gray-scale value of image, adopt the time window of T=[0,255], after coding, be converted into neuronic pulse firing sequential, particularly, for a gray value vectors (x ₁... .., x _n), wherein x _j∈ [0,255], can be encoded into n t the neuronic duration of ignition _j, wherein t _j=T-c × x _j, T represents time window, and c represents time constant, x _jrepresent input original image gray-scale value, t _jthe duration of ignition after presentation code; Time can be defined as some other relative spike sequences of same neuron generation or the triggering of multiple neuronic single stimulation, if for each neuron, think stimulate trigger after only have the delay of first spike, so just can obtain the coded system based on spike duration of ignition first time in method, adopt formula t _j=T-c × x _j, obtain the concrete Spiking pulse firing time by gray-scale value, when input picture gray scale stimulates larger, so its corresponding Spiking neuron firing intensity is stronger, and the pulse firing time of performance more early;

(33) input layer in Spiking ?convolutional network model selects the two-dimensional matrix with image dimension formed objects, each expression Spiking neuron in matrix, forms input Spiking time of its input layer according to encode each Spiking neuron firing time of obtaining of the time delay of step (32).

(4) the Spiking ?convolution algorithm based on operator is applied to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the edge of reconstruct output image.Spiking ?convolution algorithm based on operator applied to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the concrete steps at the edge of reconstruct output image are as follows:

(41) the Spiking-convolution algorithm based on operator is acted on the Spiking neuron of input layer, namely in input layer Spiking neuron, using the convolution kernel of DOG/LOG wave filter as Spiking-convolutional network, carry out Spiking-convolution, to the generation of Spiking-convolutional layer, need the smoothing filter that identical with convolution kernel size, act in the receptive field of characteristic pattern with identical weights and bias term, make the weights in its wave filter, can enter the adjustment of network synaptic plasticity, its concrete Convolution Formula is: $X_j^k=f(\underset{i\∈M_j}{\Σ}T\left(X_i^{k-1}\right)*{\mathrm{Kernel}}_{ij}^k+B^k),$ the value of Spiking-convolutional layer, be the value of input layer, T is Spiking neuron coded system, and k represents which layer of network, and Kernel is convolution kernel, and i, j represent the corresponding network number of plies, i=1 in the present invention, j=2, and each characteristic pattern can be produced by different convolution kernels, and M _jfor a selection of input feature vector figure, every one deck has unique skew B, and f activation function is Spiking pulse voltage function ${\mathrm{\ϵ}}_{ij}\left(t\right)=\frac{t-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}}\exp (1-\frac{t-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}})H(x-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}),$ ε _ijt (), as f activation function, τ is time constant, axonal delay, step function, be the input layer duration of ignition, t is time shaft, and α x is aixs cylinder constant;

(42) after Spiking-convolution, produce the neuronic voltage curve of each post-synapse Spiking immediately, namely have employed the threshold value fire model of Spiking, the neuronic mode voltage curve state of simulation biological brain, first establishes the neuronic current potential ε of all post-synapses _ijall the same, and independent of the time parameter triggered for the first time only relevant with the presynaptic signal pulse be currently received; Next makes kernel function κ also independent of the time parameter that first time triggers, and wherein the state equation of neuronal voltage is u _j(t), ε ₀s () is the PSP postsynaptic function of voltage simplified, formula is: wherein τ is time constant, axonal delay, be step function, s is time shaft, and α x is aixs cylinder constant, and the present invention have ignored the effect after resting potential, namely sets its model each optimum configurations is in the present invention: timeconstantτ=2ms, transmission delay Δ=4ms in aixs cylinder, ignition threshold value θ=0.08mv, material is thus formed the neuronic voltage curve of Spiking in Spiking-convolutional layer, is namely the method that whole threshold value fire model adopts;

(43) according to the threshold value fire model (SRM) of Spiking, setting threshold value, the neuronic voltage curve of record post-synapse Spiking arrives the precise time point of threshold value, thus create the neuronic igniting sequential of Spiking post-synapse, wherein there are kind of special circumstances, namely be Spiking neuron in whole time window, when its maximum magnitude of voltage does not arrive threshold value, recording its Spiking neuronic duration of ignition is 0; Namely after Spiking neuron accepts voltage input, because function of voltage ε _ijt () is that postsynaptic voltage (PSP) is with burst length continually varying curve, so will as the output of discrete picture, we use Spiking neuron pulse precise time treatment mechanism, the pulse firing time of postsynaptic neuron is exported as image digitization, namely by formula and record the pulse firing time as the neuronic output valve of convolutional layer Spiking.The Spiking threshold value fire model adopted in the present invention, setting threshold value θ, record convolutional layer voltage arrives the precise time point of θ, produces Spiking post-synapse neuronic duration of ignition; If the Spiking neuron wherein had is in whole time window, if maximum voltage value does not arrive threshold value θ, the duration of ignition of recording its post-synapse is 0;

(44) the igniting sequential of Spiking post-synapse neuron generation, as the numerical point of the Spiking-convolutional layer in Spiking-convolutional neural networks;

(45) step (44) is passed through decoding and reconstituting, namely be directly as image intensity value using neuronic for the Spiking post-synapse duration of ignition, namely direct by the duration of ignition in the Spiking time window recorded, as image gray-scale value and image is normalized, be between window [0,1] by pixel value normalizing;

(46) after image being normalized, carry out binaryzation operation to output layer image, producing output layer, is namely the marginal information of image, to [0,1] pixel value in data, image digitization two-value, as threshold value, is turned to 0 and 1 by peek value 0.45, the image information exported is marginal information, in sum, whole Spiking ?convolution mechanism, edge detection process is as shown in figure 11.

The present invention is illustrated by above-described embodiment, but should be understood that, above-described embodiment just for the object of illustrating and illustrate, and is not intended to the present invention to be limited in described scope of embodiments.In addition it will be appreciated by persons skilled in the art that the present invention is not limited to above-mentioned exemplifying embodiment, more kinds of variants and modifications can also be made according to instruction of the present invention, within these variants and modifications all drop on the present invention's scope required for protection.Protection scope of the present invention defined by the appended claims and equivalent scope thereof.

Claims (6)

1., based on a method for detecting image edge for Spiking ?convolutional network model, it is characterized in that, following steps:

(1) the information processing connected mode of view-based access control model hierarchy, be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?convolutional network model;

(2) the Spiking-convolutional network model of convolutional coding structure will built, use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?convolution algorithm;

(3) obtaining image, is Spiking neuron by image intensity value pixel coder, as Spiking ?the input layer of convolutional network model;

(4) the Spiking ?convolution algorithm based on operator is applied to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the edge of reconstruct output image.

2. according to claim 1 a kind of based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, in described step (1), be built with input layer, Spiking ?convolutional layer and output layer convolutional coding structure Spiking ?the concrete steps of convolutional network model as follows:

(11) set up one " Shu Ru Ceng ?Spiking ?Juan Ji Ceng ?output layer " pattern 3 layers of Spiking ?convolutional neural networks structure;

(12) according to Spiking ?the data characteristics of convolutional network structure and two dimensional image, set the two-dimensional matrix of and pretreatment image identical dimensional, and make each pixel be mapped to one by one Spiking ?input layer in convolutional neural networks structure;

(13) Spiking ?in convolutional network structure, the vision system that simulation is biological, the function of emulation receptive field, by Spiking ?input layer in convolutional neural networks structure to Spiking ?Spiking in convolutional neural networks structure ?the connected mode of convolutional layer be reduced to subregion and connect, obtain the convolution partially connected mode of Spiking nerve impulse;

(14) after obtaining the convolution partially connected mode of Spiking nerve impulse, again to Spiking ?in convolutional neural networks structure Spiking ?each receptive field of convolutional layer adopt weights to share, the weights that all receptive fields adopt are all identical, namely the wave filter of each Spiking-convolutional layer can repeatedly act in each region, the result of input signal being carried out to convolution constitutes input signal feature, thus extract the local feature of signal, the wave filter of each Spiking-convolutional layer is identical, namely identical parameter is shared, comprise identical weight matrix and bias term, final generation Spiking ?convolutional network model.

3. according to claim 1 a kind of based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, in described step (2), use Laplce's Gauss operator (LOG) and difference of Gaussian (DOG) respectively as the wave filter of Spiking-convolutional layer, formed based on operator Spiking ?the concrete steps of convolution algorithm as follows:

(21) according to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Laplce's Gaussian filter;

(22) according to difference of Gaussian (DOG) function, difference of Gaussian (DOG) functional form is converted into the gradient template identical with convolution kernel size, and there is the equal effect of Difference of Gaussian filter;

(23) by the gradient template in step (21) and step (22), as the convolution kernel in Spiking-convolutional network model, with identical template repeat function in each receptive field region, form the Spiking-convolution algorithm based on operator.

4. according to claim 3 a kind of based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, in described step (21), according to Laplce's Gauss operator (LOG) function, Laplce's Gauss operator (LOG) functional form is converted into the gradient template identical with convolution kernel size, and the concrete steps with the equal effect of Laplce's Gaussian filter are:

(211) first with Gaussian function to the smoothing filtering of image;

(212) smoothing filtered image is carried out Laplace computing;

(213) after will carrying out Laplace computing, null point is as marginal point.

5. according to claim 1 a kind of based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, in described step (3), be Spiking neuron by image intensity value pixel coder, as Spiking ?the concrete steps of input layer of convolutional network model as follows:

(31) by the pixel of input picture, image gray processing process is carried out, in pixel value range is limited to [0,255] interval;

(32) Spiking time delay encoding operation is carried out to each gray-scale value of image, adopt the time window of T=[0,255], adopt formula t _j=T-c × x _j, T represents time window, and c represents time constant, x _jrepresent input original image gray-scale value, t _jthe duration of ignition after presentation code, obtain the concrete Spiking pulse firing time by gray-scale value, when input picture gray scale stimulates larger, so its corresponding Spiking neuron firing intensity is stronger, and the pulse firing time of performance more early;

(33) input layer in Spiking-convolutional network model selects the two-dimensional matrix with image dimension formed objects, each expression Spiking neuron in matrix, forms input Spiking time of its input layer according to encode each Spiking neuron firing time of obtaining of the time delay of step (32).

6. according to claim 5 a kind of based on Spiking ?the method for detecting image edge of convolutional network model, it is characterized in that, in described step (4) by based on operator Spiking ?convolution algorithm apply to Spiking ?convolutional network model, pulse convolution is taked to input layer, and then according to Spiking threshold value fire model, the concrete steps at the edge of reconstruct output image are as follows:

(41) the Spiking-convolution algorithm based on operator is acted on the Spiking neuron of input layer, namely in input layer Spiking neuron, using the convolution kernel of DOG/LOG wave filter as Spiking-convolutional network, carry out Spiking-convolution, its concrete Convolution Formula is: $X_j^k=f\left(\underset{i\∈M_j}{\Σ}T\right(X_i^{k-1})*{\mathrm{Kernel}}_{ij}^k+B^k),$ the value of Spiking-convolutional layer, be the value of input layer, T is Spiking neuron coded system, and k represents which layer of network, and Kernel is convolution kernel, and i, j represent the corresponding network number of plies, and each characteristic pattern can be produced by different convolution kernels, and M _jfor a selection of input feature vector figure, every one deck has unique skew B, and f activation function is Spiking pulse voltage function ${\mathrm{\ϵ}}_{ij}\left(t\right)=\frac{t-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}}\exp (1-\frac{t-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}}{\mathrm{\τ}})H(x-t_i^f-{\mathrm{\Δ}}_{ij}^{ax}),$ ε _ijt (), as f activation function, wherein τ is time constant, axonal delay, step function, be the input layer duration of ignition, t is time shaft, and α x is aixs cylinder constant;

(42), after Spiking-convolution, the neuronic voltage curve of each post-synapse Spiking is produced immediately;

(43) according to the threshold value fire model SRM of Spiking, setting threshold value, the neuronic voltage curve of record post-synapse Spiking arrives the precise time point of threshold value, thus create the neuronic igniting sequential of Spiking post-synapse, wherein there are kind of special circumstances, namely be Spiking neuron in whole time window, when its maximum magnitude of voltage does not arrive threshold value, recording its Spiking neuronic duration of ignition is 0;

(44) the igniting sequential of Spiking post-synapse neuron generation, as the numerical point of the Spiking-convolutional layer in Spiking-convolutional neural networks;

(45) by step (44) by decoding and reconstituting, Spiking post-synapse neuronic duration of ignition directly as image intensity value, and is normalized image;

(46) after being normalized by image, carry out binaryzation operation to output layer image, producing output layer, is namely the marginal information of image.