Study On The Methods For Image Storage And Retrieval Modelling Based On Human Brain Memory Mechanism

With the rapid development of computer technology and network multimedia technology,a huge amount of image data is generated every moment.How to store and retrive these image data efficiently is a key technical problem to be solved in computer vision.As we all know,human visual system collects a lot of visual information every day.Despite the complexity of the external world,the human brain is able to store it effectively and recall it quickly when needed.Therefore,applying the human brain memory mechanism to computer vision and image processing has significant academic and applicable value.This paper maily studies the human brain memory mechanism on the basis of research results from cognitive psychology and neuroscience and applies it to visal image classification memory.The main work of this paper is as follows:1.The related theories of human brain memory are studied,including the associative memory model based on neurophysiology,the free energy minimization theory based on Bayesian brain and the probabilistic generative modelling process of restricted Boltzmann machine.2.An incremental pattern association memory model is proposed,in which a Leabra learning framework based on error-driven and Hebbian associative learning is adopted to realize incremental image storage and fast retrival.Compared with the other three models on four standard image databases,the results show that the proposed incremental pattern association memory model has better image classification performance and higher time efficiency.3.A classification restricted Boltzmann machine memory model based on free energy minimization is proposed.The theory of free energy proposed by Friston is introduced into the restricted Boltzmann machines to optimize the memory model parameters,in which the posterior estimation is maximized to ensure a good approximation to the true recognition density,and the input likelihood is maximized to enable the generative model to match the input distribution well.Experimental results on image classification show that the performance of the proposed memory model is better than the existing self-organizing incremental neural network and support vector machine,and it can recall the visual image memory as well.