Research And Application Of Gabor Convolutional Neural Network Model

Convolutional neural network(CNN)is a classic artificial neural network model.It is one of the artificial neural networks with superior performance in the field of deep learning.It has been widely used in the field of computer vision,such as image classification and target Recognition,image segmentation,semantic recognition,etc.The development of convolutional neural networks has greatly promoted the research and promotion of modern technologies such as artificial intelligence,which has farreaching significance.The good expressiveness of the convolutional neural network model depends deeply on the deep-level structure of the network model,which is conducive to extracting more feature information.However,the deepening of the network model means an increase of convolution kernels,which leads to a large number of training parameters during the network learning,occupying memory and consuming computing resources;In addition,since the convolution kernel does not possess the directional characteristics and spatiotemporal frequency characteristics of the visual cell receptive field,using the convolution kernel as the receptive field cannot completely simulate the human visual perception mechanism.Aiming at the shortcomings of convolutional neural networks,a Gabor convolutional neural network calculation model is proposed in this paper that simulates the characteristics of visual neurons.The innovations of the model are shown in the following:Firstly,based on the calculation model of the two-dimensional Gabor convolution kernel,a two-dimensional Gabor convolutional neural network is constructed.The twodimensional Gabor filter is used to meet the visual perception characteristics of human visual primary cortical cells,and the proposed two-dimensional Gabor convolution kernel is used to simulate the human visual receptive field to obtain direction information and time-frequency domain information;A two-dimensional Gabor convolutional neural network is established,discussed the Gabor filter parameters(such as direction,frequency,etc.)which reflect the visual attributes through the learning method of backpropagation,so as to realize the analysis of the image spatial information and the recognition of different objects in the image.Secondly,the calculation model of the 3D Gabor convolution kernel is proposed,and the 3D Gabor convolution neural network based on the 3D Gabor convolution kernel is established.On the basis of the two-dimensional Gabor filter,the network establishes a three-dimensional Gabor filter based on the inseparable specificity of the time and space of the visual processing cortical neurons,analyzes the role of the filter parameters,and uses the three-dimensional Gabor convolution kernel to construct the threedimensional Gabor Convolutional neural network analyzes the video sequence information,extracts spatiotemporal information and optical flow information,so as to realize the action recognition in the amount of video data.Finally,a neural network experimental system based on two-dimensional and threedimensional Gabor convolution kernels is established.Based on the two-dimensional Gabor convolutional neural network and the three-dimensional Gabor convolutional neural network,the image classification experimental system and the action recognition experimental system were established respectively,and tested on the public image data set CIFAR10,Cat&Dog,Mini-imagenet and the action recognition data set KTH.The experimental test results show that the proposed network model proposed in this paper has better network performance.In this paper,the research on the combination of Gabor filter and convolutional neural network is based on the relevant theoretical basis.The experimental verification shows that Gabor filter has application value in both two-dimensional and threedimensional levels.Combining Gabor filter with the convolutional neural network with superior performance can complement each other,thus reflecting the advantages of both.Therefore,the proposed network model proposed in this paper has strong validity and integrity.