Research And Implementation Of Image Recognition Based On Spiking Neural Network

Deep learning has achieved a series of breakthrough performances in the field of image recognition.However,this complex deep neural network model requires huge calculation and storage requirements,which causes great difficulties for them to implement on platforms with limited resources.Impulsive neural networks are inspired by spiking neuromorphic calculations,and are generally considered to be the third generation neural networks that can closely link the working mechanism and efficiency of the human brain with deep learning,which can achieve greater computing efficiency at lower power and can adapt to different resource platforms.The most important thing in the impulsive neural network is the spiking signal,which is to simulate brain neurons to encode neural information using event-driven signals.spiking neural networks use this advantage as a biologically relevant model.However,the current research is limited by the inability to effectively encode external stimulus information and the shallow network structure.In order to solve this limitation,this paper proposes a novel hybrid framework by being inspired by the asymmetric processing of low frequency spatial information and high frequency spatial information in the left and right hemispheres of the human brain.This new framework combines a neural network(DenseNet)with excellent feature extraction capabilities and a spiking neural network.Impulsive neural networks can be guided by DenseNet in the process of feature extraction to learn powerful feature extraction capabilities.This paper applies this method to the field of image recognition.In terms of model accuracy and robustness,this kind of network shows good performance in image recognition.The work of my thesis is as follows:1)My thesis analyzes the current neural information coding and proposes a coding method based on time frequency in spiking neurons.The coding error tolerance rate and accuracy designed by this coding method are improved.2)My thesis studies and analyzes the current impulsive neural network model architecture,and proposes a new hybrid neural network architecture.The spiking neural network is guided by DenseNet to make up for the defect feature extraction ability.At the same time,the use of rich feature information through the characteristics of the impulsive neural network enhances the understanding of the image,and improves the accuracy,robustness and robustness of the network for image recognition.3)For the designed new impulsive neural network,the ability to extract features needs to be guided by DenseNet,so on this basis,a gated network based on Tempotron supervised learning algorithm is designed as a connection guidance control...