Study On The Methods For Visual Image Classification Based On Associative Memory Neural Network

Visual image classification is the foundation for computer vision and pattern recognition,and has become a hot research topic for a long time.Although a large number of methods have been proposed to date,these methods still can’t understand and classify complex visual images as quickly as human beings.The reason why human can effectively represent,store and recognize massive visual information is that the human brain can associate different feature information of the same thing to form associative memory and complete the information using part of features.Therefore,it is of great theoretical and practical values to study the associative memory mechanism of human brain and apply it to visual image classification.To this end,this paper mainly studies the associative memory neural network based on the research results on human brain memory from neurobiology and cognitive neuroscience,then applies it to visual images association and classification.The main work of this paper is as follows:1.The neurobiological basis of associative memory in human brain and the encoding,storage and retrieval process of memory information are studied,then the structure and modeling process of the three typical associative memory neural networks are summarized,which lays a theoretical foundation for the application of associative memory mechanism in visual image classification.2.A visual image classification algorithm based on auto-associative memory neural network is proposed,which uses Hebb-like local association learning rules to store the association information between image features and their categories in the synaptic connection of the network,thus breaking through the limitation that the original model can only process binary information.Image classification experiments are first carried out on four benchmark image databases,and then compared with support vector machine(SVM),back propagation neural network(BPNN)and other associative memory models.The results validate the effectiveness of the proposed algorithm.In addition,the proposed model is also applied to pattern completion using an image dataset consisting of ten capital letters.3.A visual image classification algorithm based on bidirectional pattern association memory neural network is proposed.The improved time difference Hebbian association learning rule and k WTA(k-winners-take-all)function are introduced into Hebbian learning mechanism,which overcomes the limitation that the original model can only transmit in one direction and the disadvantage that the weight may increase unlimitedly in Hebbian learning,so that the network can learn the information of paired images as well as the association information between image features and their category label encodings.Experiments on ten pairs of “face-fashion” images show that the proposed algorithm is able to achieve association memory between images.Furthermore,experiments on three multi-classification databases show that the proposed algorithm has higher classification accuracy than SVM,BPNN and other associative memory models.