Research On Image Classification Algorithm Of Spiking Neural Network Based On Multi-scale Receptive Field

Artificial intelligence aims to simulate and achieve certain human intelligent behaviors through computers.Vision is the most important medium for humans to obtain information from the outside world.Computer vision is a technology to complete visual input and understanding through the computer instead of humans.Image classification,a basic task of computer vision,is of great significance to simulating and achieving human visual systems.However,the traditional artificial neural networks based on BP(Error BackPropagation)have some problems in simulating human intelligence,such as high computational power consumption and high requirements for the hardware equipment.In this context,the Spiking Neural Network has attracted more and more researchers’ attention due to its low power consumption,event-driven characteristics and high bio-interpretability.However,the current SNN image network has problems such as the structure being single or the models lacking biological characteristics.Therefore,this paper proposes a series of image classification algorithms with bio-interpretability and small network scale by simulating the human visual system.It is mainly divided into the following contents:(1)Design a locally connected multi-scale feature learning network.Inspired by the receptive field distribution of retinal cells,a network structure with local connections is designed and feature learning is carried out at different scales,indicating that feature learning networks of different scales can learn different features.(2)Study the image classification algorithm based on receptive field sliding.The input image is spatially divided into several local inputs and the locally connected network is used to learn local features.The learning process of receptive field sliding is designed and a supervised learning rule with biological rationality is adopted to integrate multiple local features in time.The classification accuracy is 85.34% on the MNIST dataset.Thanks to the local connection,the algorithm achieves the same classification accuracy as the fully connected network with smaller network size and fewer training parameters.(3)Study the image classification algorithm based on the multi-scale receptive field.A multi-scale receptive field with a "dense and small center and sparse and large periphery" was designed,and a locally connected network was used to learn multiple local spatial features.Three classification methods are proposed to integrate multiple local features in space.This algorithm has high bio-interpretability and good classification effect,and finally achieves 96.91% classification accuracy on the MNIST dataset.