Research On Image Recognition Based On Spiking Neural Network

In recent years,while moving ahead with big data technology and parallel computing technology,deep networks attempt obligated significant breakthroughs in image recognition tasks.However,the neurons of this kind of deep network are based on real-valued calculations,using back-propagation for network learning,and as the network depth continues to increase,requires significant computing and memory resources,which limits the practical application of deep networks development.A new modern artificial neural network inspired by brain science,the spiking neural network deeply simulates the dynamic properties of biological neurons and is currently the most biologically plausible neural network.Event-driven spiking neural networks provide sparse and powerful computing power,which is expected to solve the limitations of deep networks of practical applications.Therefore,how to integrate the advantages of the spiking neural network and the convolutional neural network and design a network model with high computational efficiency and high accuracy has important practical significance.Imaginative by the mechanism of biological vision,this thesis conducts an intensive study on the problem of visual information spiking coding and convolutional spiking neural network image recognition and the concrete research responsibility as follows:(1)Visual feature coding based on time series.According to the characteristics of the receptive field of the retinal ganglion cells on the biological vision pathway,a Gaussian difference filter is used to simulate the characteristics of the receptive field to extract the edge features of the image from the input information.In order to fix the information extracted by the single-scale Gaussian difference filter,a multi-scale Gaussian difference filter is designed to simulate the characteristics of the multi-scale receptive field with different window sizes.Combining with the theory of spiking timing coding,a spiking coding method based on edge features is proposed,which encodes the extracted image edge features into spiking time series.The application of the Gaussian difference filter extracts rich image edge features and retains the contour information important for image recognition.(2)Convolutional spiking neural network image recognition based on the attention mechanism.Inspired by the hierarchical information processing mechanism of the visual system and the spiking information representation and transmission mechanism of the visual system,combined with the time-series-based single-scale receptive field visual nerve information coding method in(1),it integrates the spiking neural network and the convolutional neural network.Advantages,proposed a convolutional spiking neural network image recognition model based on the attention mechanism,referred to as AMCSNN.The model is based on the unsupervised spiking timing-dependent plasticity learning rules of synaptic plasticity to learn and update the network weights.It focuses on the important information features of the image by combining the lightweight channel space attention mechanism and uses the additional supervision module SVM to complete the image recognition task.The experimental analysis of the designed network model on multiple image recognition benchmark data sets shows that the network model achieves excellent recognition accuracy while also achieving faster network calculation efficiency,the validity of the proposed impulsive neural network model was verified.(3)Convolutional spiking neural network image recognition based on reward and punishment mechanism.Aiming at the problem that the support vector machine classifier used in the network model in(2)has no reasonable biological practical significance and poor bionics,it is inspired by the brain reward and punishment mechanism and combined with the time series-based multi-scale receptive field vision proposed in(1)Neural information coding method,proposed a convolutional spiking neural network image recognition model based on reward and punishment mechanism,referred to as RPCSNN.This model simulates the decision-making behavior of the brain based on the reward and punishment mechanism and has stronger biological authenticity.The experimental analysis of the designed network model on several image recognition benchmark data sets shows that the network model has high recognition accuracy,strong network computing ability,and better bionics,which has a certain reference significance for the representation of brain information and the simulation of biological mechanism.