Research Of Pattern Recognition Based On Spiking Neural Networks

In the ability to identify objects,the biological vision system has more powerful capabilities than current computer vision systems.According to the latest neuroscience research results,if we use proper coding scheme and supervised learning rules to establish a spiking neural network,we can use this network to simulate the biological nervous system.Through encoding,external stimuli is converting into sparse representations;through the learning algorithm,we can learn the internal laws contained in these spiking signals.In this paper,we use proper coding scheme and supervised learning rules to establish a spiking neural network based on the distinction and invariability of visual recognition.In addition,we also put forward an improved Tempotron learning rule with the aim to increase its robustness even under noise condition.The main contributions of this article are as follows:(1)We put forward a method to extraction image feature.This method can simulate biological vision system and has a good feature representation.According to the similarity between the Gabor function and the receptive field of human visual neuron cells,this method extracts the image features by Gabor filter convolution operation at multiple scales and multiple angles.For the problem of convolutional layer information redundancy,this paper studies the characteristics of Gabor filters,and proposes a convolution layer sparse method based on the gradient direction of the edge lines.Finally,according to the characteristics of spiking neural network and the theory of delay coding,this paper improves the pooling method.Convert features to spiking sequences.The features take into account the distinction and invariability,and the important information of the image is preserved.(2)This article improves the classic learning algorithm Tempotron.The improved algorithm analyses the Tempotron's two misclassifications,and add the correction item to control the spike time.Experimental results demonstrate that the improved Tempotron learning rule can achieve a high classification accuracy even under noise-condition.(3)A recognition model based on spiking neural network for biological vision simulation is proposed.With a combination of the image feature extraction method of(1)and the improved Tempotron learning algorithm of(2),the proposed computational model can extract image features efficiently and encodes them as spike signals,and then uses a learning algorithm to train the network.Finally,was test the effect of the model in the Caltech 101 dataset.The results show that the pattern recognition model presented in this thesis is equivalent to the RFCS model in terms of recognition accuracy;in the environment of noise interference,the recognition accuracy of this model is slightly better than the RFCS model[22].In general,the model proposed in this paper has the advantages of good bionics performance,high recognition performance and strong robustness.These results also provide reasonable proofs for the simulation of biometrics.