Research On Neural Networks With Visual Functions For Color Image Classification

With the maturity of color imaging technology and the popularization of color images,color information has received more and more attention in computer vision tasks.However,color is an optical characteristic.The color feature extracted by the existing algorithm is not robust.In order to have the same excellent color information processing ability as human eyes for computer vision,it is necessary to start from human vision functions and make full use of the powerful feature extraction ability of artificial neural network.The wave of artificial intelligence has swept the world,and the artificial intelligence companies are proliferating in recent years.Bionic vision is the key to break through the bottleneck of artificial intelligence development.The combination of visual system and neural network is also a very important and interesting research topic.However,the existing bionic visual neural network models did not fully simulate the complex functions of human visual system.More appropriate visual fuctions need to be introduced in order to improve the performance of artificial neural network comparable to that of human vision.In addition,most of the traditional neural network models are usually oriented towards fixed visual tasks or single structure neural networks,which are difficult to adapt to the complex task needs.The application development of general artificial intelligence in the future needs to combine the advantages of various types of neural networks,and realize the fusion and switch between different tasks and different networks by constructing heterogeneous neural networks.Therefore,the research and construction of different types of bionic vision neural network model are the basis of building heterogeneous neural network research and urgent demand.Based on the processing process of color and other information of human vision,this paper constructs and simulates the neural network model of retina,primary visual pathways and middle and high-level visual pathways according to the hierarchical structure of human visual pathways.The different hierarchical models are used in different classification and recognition tasks of color image sets,which lay a foundation for heterogeneous neural network fusion in the future.Research contents are as follows:First,the multi-channel convolution neural network model was constructed based on the retinal perception characteristics such as color adaptability,brightness adaptability and color constancy.Also,it combined the two-stage vision theory model with the convolution neural network model.The model reduces the influence of external light source on useful information by simulating the pre-processing operation of human vision system.According to opponent color and bright vision mechanism,the vision information is recoded to enhance the color representation.The results show that the model is suitable for the classification of scenes with large change of light source or low light,and has the characteristics of color constancy.By comparing the strategy of deepening the layer number of convolutional neural network,the method of combining retinal vision function and convolutional neural network can also get the same classification accuracy.Second,based on the analysis of primary visual pathways and the characteristics of single and double opponent fields of neurons,a neural network model based on the information processing hierarchy of human visual system is constructed.In this network the neurons with a receptive field in center-surround concentric circle are trained using Spike Timing Dependent plasticity(STDP)learning rules.In this way,the performance of feature extraction and classification are improved.The results show that this method is suitable for the classification of objects with rich color and obvious color contrast.Compared with other methods,this model achieves higher classification accuracy.Finally,a CNN model is designed based on the characteristics of single opponent neurons and the higher level attention mechanism of visual channel.On the basis of the two-stage theory model,the paper uses a mechanism for transform of the visual attention focus to search the discriminative local region and carry out the opponent color space conversion.The convolution layer of CNN is used to extract the local key features and overall features.Then,all feature components are combined together for training of the classification network.The research results show that the model can search out the detailed features of objects well,which not only reduces the complexity and workload of annotation,but also achieves better results by comparing other fine-grained methods.