Study On Object Recognition Models Based On Visual Pathway

Object recognition base on Primate Visual Cortex is the current hotspot in computer area,which using various aspects of neuron fire experiment to establish the model of image processing. Object recognition and computer vision both have the aim of studying and embodying the visual cognitive ability of human, thus the research on image processing from the perspective of human visual perception system has important theoretical sigificance and application prospect. How using a algorithm to simulate visual perception, a object recognition model based on visual pathway which inlcuding retina, LGN and Primate Visual Cortex was established in this paper.Including the following contents in this dissertation:(1) Convolutional Neural Network is introduced. The thought of how to building the overall Convolutional Neural Network structure is analysed through the learning of mechanism and algorithms of simple cell and Complex Cell in Hidden Layer, also the superiority of learning rule in Hidden Layer is summarized.(2) Sparse Coding is analyzed. The objective function and the learning rule are explained, that shows the application through feature extraction using Gabor filter to initialize the Sparse Coding primary function(3) The physiological meaning and arithmetic in each layer of HMAX model are discussed. Then original max Pooling is compared with average Pooling through data classification experiment, which verifies advantage of max pooling so as to lay a theoretical foundation for model selection.(4) An algorithm, four layers model each contain coding and pooling step using convolutional neural network structure is established base on complete visual pathway. In coding step, a nonlinear filter which consistent with neurons from retina to the V1 in visual cortex is designed from the gradient descent algorithm in sparse coding.In pooling step, max Pooling in HMAX model which consistent with neurons from V2 toV4 in visual cortex was adapt. The experiment results show that the method can effectively reduce algorithm complexity and improve accuracy of classification.