CN108846349A - A kind of face identification method based on dynamic Spiking neural network - Google Patents

Abstract

The invention discloses a face recognition method based on a dynamic Spiking neural network, and relates to the technical field of image processing. The invention comprises the following steps: S1, converting a face image into grayscale pixels to obtain a grayscale image; S2, grayscale feature extraction from the degree map to obtain the low-dimensional features associated with the regional features of the gray-scale image; S3, convert the feature intensity of the low-dimensional features into a pulse time series; S4, train the Spiking neural network with multiple face images, and obtain the dynamic Spiking Neural network; S5, the pulse time series obtained after the face image to be recognized is processed by S1-S3 is input into the dynamic Spiking neural network obtained by S4, according to the adjustment weight and the existing neuron in the dynamic Spiking neural network Weights are compared, and the category of the neuron with the closest weight is the category of the face image to be recognized. The present invention creatively uses the dynamic Spiking neural network, which significantly improves the recognition efficiency compared with the traditional Spiking image recognition method. .

Description

A Face Recognition Method Based on Dynamic Spiking Neural Network

technical field

The invention relates to the technical field of image processing, and more specifically relates to a face recognition method based on a dynamic Spiking neural network.

Background technique

Face recognition is a new research direction based on graphics, computer science and technology, pattern recognition and other related disciplines. When distinguishing human beings by extracting human characteristics, it is the most concise way to use human faces as the basis. The way. Although face recognition technology has only been around for decades, it has become one of the more popular research topics at present. Especially in the era of artificial intelligence, with the rapid development of science and technology and people's pursuit of a safe and intelligent life, the requirements for the effectiveness, convenience, and speed of face recognition technology are getting higher and higher.

As a third-generation neural network, the Spiking neural network focuses on the timing of pulses and is suitable for implementation on a chip. However, most supervised or unsupervised learning algorithms for spiking neural networks have a fixed structure, in which the size of the hidden layer and output layer must be specified in advance, and are trained in offline batch mode. Therefore, these methods can only be applied to class or the number of clusters is known; moreover, these methods cannot be applied to problems with continuously changing data, since they will require retraining on old and new data samples. However, biological neural networks are well known for their ability to learn continuously and incrementally, which enables them to continuously adapt to changing non-stationary environments, therefore, in order to allow SNNs (SpikingNeuron Networks, Spiking Neural Networks) to interact with continuously changing environments , it is necessary to adapt its structure and weights dynamically to new data, moreover, when learning new information, catastrophic interference or forgetting should be avoided.

Contents of the invention

The purpose of the present invention is: in order to solve the prior art problem that face recognition is highly dependent on the category of samples, the present invention provides a face recognition method based on dynamic Spiking neural network.

The present invention specifically adopts the following technical solutions in order to achieve the above object:

A face recognition method based on dynamic Spiking neural network, comprising the following steps:

S1, converting the face image into grayscale pixels to obtain a grayscale image;

S2. Perform feature extraction on the grayscale image to obtain low-dimensional features associated with regional features of the grayscale image;

S3, converting the feature intensity of the low-dimensional feature into a pulse time series;

S4, using multiple face images to train the Spiking neural network

After multiple face images are processed through S1-S3 in sequence, the corresponding pulse time series are respectively obtained, and the corresponding initial weights are adjusted according to the precise time of the pulses in each pulse time series, and the adjusted weights are obtained, and then the weights are carried out. Value learning, according to the label of each face image, adjusting the weight and the existing neurons in the Spiking neural network to jointly judge whether to add new neurons, when all the face images are input into the Spiking neural network, get Stable dynamic Spiking neural network;

S5. Recognize the face image to be recognized

Input the pulse time series obtained after processing the face image to be recognized through S1-S3 into the dynamic Spiking neural network obtained in S4, adjust the initial weight according to the precise time of the pulse in the input pulse time series, and get the corresponding adjustment Weight, compare the adjusted weight with the weight of the existing neurons in the dynamic Spiking neural network, find out the neuron with the closest weight to the adjusted weight, and the category of the neuron is to be identified The category of face images.

Further, in the above S2, feature extraction is performed on the grayscale image using the PCA dimensionality reduction method.

Further, in S3, the Gaussian encoding method is used to convert the feature intensity of the low-dimensional feature into a pulse time series.

Further, the precise time-based weight learning in S4 is realized by using a decreasing function, that is, the first arriving pulse contains the most information.

The beneficial effects of the present invention are as follows:

1. The present invention extracts the features of the face image to be recognized based on the idea of PCA dimensionality reduction, and then uses the Gaussian encoding method to encode the extracted features, and uses the dynamic Spiking neural network learning algorithm to encode the encoded features after encoding. Sequential learning, adjust the initial synaptic weights of the input pulse time series according to the precise time of the pulse, dynamically increase or update neurons according to the similarity between the weights, and finally obtain the recognition output result, creatively using dynamic Spiking Compared with the traditional Spiking image recognition method, the neural network has significantly improved the recognition efficiency.

2. The present invention adjusts the initial synaptic weights of the input pulse time series according to the precise time of the pulse, which improves the accuracy of face image recognition.

3. The present invention uses similarity to compare weights, which further improves the efficiency of face image recognition.

4. The present invention uses a dynamic Spiking neural network structure to dynamically increase neurons according to the input samples, and does not need to make statistics on the labels of the samples in advance, which reduces the dependence of the face recognition method on the samples.

Description of drawings

Fig. 1 is a schematic diagram of the identification process of the present invention.

Fig. 2 is a schematic diagram of the overall network structure of the present invention.

Fig. 3 is the training flowchart of the dynamic Spiking neural network of the present invention.

Figure 4 is a sample diagram of the ORL dataset.

Detailed ways

In order for those skilled in the art to better understand the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and the following embodiments.

Example 1

As shown in Figures 1 to 4, the present embodiment provides a face recognition method based on a dynamic Spiking neural network, comprising the following steps:

S1, converting the face image into grayscale pixels to obtain a grayscale image;

S2. Using the method of PCA dimensionality reduction to perform feature extraction on the grayscale image to obtain low-dimensional features associated with regional features of the grayscale image, including the following steps:

S2.1. Express an N-dimensional column vector as x _i , where i=1,2,...,L;

S2.2. Calculate the average value x _i of the L sample vectors;

S2.3, calculate the covariance matrix;

S2.4. Off-diagonal elements are the correlation between elements of each column vector, and the calculation formula is:

S2.5. Perform eigendecomposition on the covariance matrix to obtain several eigenvalues;

S2.6. Sorting several eigenvalues from large to small, and taking the eigenvectors corresponding to the first r eigenvalues to obtain a projection matrix;

S2.7. Use the projection matrix to calculate a new low-dimensional vector, that is, a low-dimensional feature;

S3. Using the method of Gaussian encoding to convert the characteristic intensity of the neurons in the first layer into a pulse time series, specifically:

Assuming that the low-dimensional vector obtained in S2.7 has m-dimensional features (x ₁ , x ₂ ,...,x _m ), after group coding, m*p pulse times are obtained, and the i-th feature is first calculated at j mean of receptive fields and standard deviation The specific formula is:

in, and are the minimum and maximum values of the i-th feature, respectively; β is a parameter that affects the coverage of the Gaussian receptive field by affecting the standard deviation, and by the mean and standard deviation get the Gaussian function The calculation formula is:

Using the Gaussian function As a result, the pulse time series is obtained That is, the pulse time of each input neuron is obtained, and the calculation formula is:

S4, using multiple face images to train the Spiking neural network,

Select multiple face images in the ORL face dataset and process them sequentially through S1-S3 to obtain the corresponding pulse time series, adjust the corresponding initial weights according to the precise time of the pulse in each pulse time series, and use multiple Classify with this strategy, when all face images are input to the Spiking neural network, a stable dynamic Spiking neural network is obtained;

S4.1. Initial weight fine-tuning strategy:

An independent output layer neuron represents an input pattern. For each input sample, the output layer will create a new neuron, and the weight learning of the neuron and the encoding layer is represented by a decreasing function. , the calculation formula is:

w _ij ＝w ₀ +γexp(-t _i /τ)

where w _ij is the synaptic weight between input layer neuron i and output layer neuron j, w ₀ is the initial weight, t _i is the precise time of input layer pulses, thus replacing the order of pulses, and τ is the time constant ;

S4.2. Neuron adjustment strategy:

In the training phase, the training input samples will be presented to the Spiking neural network one by one, and then the information stored in the Spiking neural network will be compared with the information carried by the input samples, which represent the functional relationship between the input features and the sample category labels, The algorithm makes each sample choose a learning strategy:

S4.2.1. Add neuron strategy: When the difference between the information in the network and the information carried by the input samples is relatively large, new neurons in the output layer are added to record new information;

S4.2.2. Merge neuron strategy: When the information of the input sample is sufficiently similar to the information stored in the existing neurons in the network, the new neuron will be merged with the most similar neuron, and the trained output neuron represents a spatiotemporal spike Clustering of patterns, merging neurons according to the similarity between neurons and predicting class labels makes it possible to achieve fast enough learning, and at the same time provide precise time for the weight vector to make the algorithm efficient enough, the merged neurons need to update the weights , add the incoming weight to the original weight to get a new weight;

S5. Recognize the face image to be recognized

Select a face image in the ORL face data set as the face image to be recognized, and input the pulse time series obtained after the face image to be recognized through S1-S3 to the dynamic Spiking neural network obtained in S4, according to The precise time of the pulse in the input pulse time series is used to adjust the initial weight, and the corresponding adjusted weight is obtained, and the adjusted weight is compared with the weight of the existing neuron in the dynamic Spiking neural network, and the adjustment is found. The neuron with the closest weight, the category of the neuron is the category of the face image to be recognized.

The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention. The scope of patent protection of the present invention is subject to the claims. Any equivalent structural changes made by using the description and accompanying drawings of the present invention, All should be included in the protection scope of the present invention in the same way.

Claims (4)

1. a face recognition method based on dynamic Spiking neural network, is characterized in that, comprises the following steps: S1, converting the face image into grayscale pixels to obtain a grayscale image; S2. Perform feature extraction on the grayscale image to obtain low-dimensional features associated with regional features of the grayscale image; S3, converting the feature intensity of the low-dimensional feature into a pulse time series; S4, using multiple face images to train the Spiking neural network After multiple face images are sequentially processed by S1-S3, the corresponding pulse time series are respectively obtained, and the corresponding initial weights are adjusted according to the precise time of the pulses in each pulse time series, and the adjusted weights are obtained, and then weighted Value learning, according to the label of each face image, adjusting the weight and the existing neurons in the Spiking neural network to jointly judge whether to add new neurons, when all the face images are input into the Spiking neural network, get Stable dynamic Spiking neural network; S5. Recognize the face image to be recognized Input the pulse time series obtained after processing the face image to be recognized through S1-S3 into the dynamic Spiking neural network obtained in S4, adjust the initial weight according to the precise time of the pulse in the input pulse time series, and get the corresponding adjustment Weight, compare the adjusted weight with the weight of the existing neurons in the dynamic Spiking neural network, find out the neuron with the closest weight to the adjusted weight, and the category of the neuron is to be identified The category of face images. 2. a kind of face recognition method based on dynamic Spiking neural network according to claim 1, is characterized in that: in described S2, adopt the method for PCA dimensionality reduction to carry out feature extraction to gray-scale image. 3. a kind of face recognition method based on dynamic Spiking neural network according to claim 1, it is characterized in that: in described S3, adopt the method for Gaussian encoding to convert the feature intensity of low-dimensional feature into pulse time series. 4. a kind of face recognition method based on dynamic Spiking neural network according to claim 1, it is characterized in that: in the described S4, based on the weight learning of accurate time, realize with a decreasing function, i.e. at first The arriving pulses contain the most information.