Visual Information Representation,Storage And Retrieval Modeling Inspired By Human Memory Mechanism

With the rapid development of computer technology and internet,a large amount of image and video data has been produced.Therefore,how to represent,store and recall these massive data of image and video in order to classify and retrieve the required data quickly is still a critical issue to be addressed in urgent need.Although the existing image classification algorithms exhibit good recognition and classification performance for multiple categories of images,most of them focus on classification ability only while ignoring the ability of recognizing new things.Moreover,the classification mechanism is quite different from human visual pattern recognition.Human can learn to remember,classify and recognize an object easily,which is intimately associated with human memory mechanism.Therefore,it is of great theoretical and applicable value to apply the human memory mechanism to computer vision to study how visual information is represented and stored in human brain,and how to retrieve it when needed.The main contributions of this dissertation are as follows.1)Firstly,the Retrieving Effectively from Memory(REM)model has been studied thoroughly.REM model was initially studied and applied in the learning and memory of words.In this work,we introduce REM memory model into the learning and memory of natural images,and propose a modified REM memory model for representing,storing and retrieving visual images.Experimental results show that REM model can be appliedfor natural image storage and retrieval,and gain good recognition effect not only in the classification of the same object with small rotation angles but also in the classification of the same category object.2)Based on the REM model,a new memory model for visual image storage and recall based on sparse coding and Bayesian decision(VIRSRBD)which is more suitable for image processing is proposed.The dense scale invariant feature transform(SIFT)features of the images are firstly extracted and represented sparsely for feature matching.Then,the likelihood value is calculated on the base of match result.Finally,the recognition and classification rule based on Bayesian decision is given.The proposed VIRSRBD model can explain the visual image classification and recognition process from the perspective of human memory model.Under the guarantee of accuracy,the false alarm rate of the model is lower than that of support vector machine(SVM),sparse representation based classification(SRC)as well as extreme learning machine(ELM).When the number of image categories is known,the classification performance of the proposed model is better than SVM method on two datasets.3)In order to improve the accuracy of the image features,the convolutional neural network(CNN)of deep learning is combined with the VIRSRBD model,and a memory model for image recognition and classification based on convolutional neural network and Bayesian decision is proposed.First the image features are extracted by convolutional neural network and then stored in binary form.Then,the matching rule and the decision criterion for classification are built according to Bayesian theory.Experimental results show that the proposed memory model for image recognition and classification can perform much better than some existing methods.With the proposed method,the recognition rate is higher than SRC and ELM,while the false rate is far lower.4)In order to simulate the visual short-term memory process of humain brain,the probabilistic clustering theory of the organization of visual short-term memory is introduced into visual image memory modeling.The latent Dirichlet allocation(LDA)model is applied to describe an image,and a novel image classification is proposed based on image semantic information.First the bag of words(BoW)model is produced by K-means algorithm,then the parameters of LDA model are estimated by Gibbs Sampling and the distribution feature of latent topic in an image is calculated.Finally,according to the classification rule based on Bayesian decision in VIRSRBD model,the visual images are classified in topic space.Experimental results show that the proposed algorithm has better performance than LDA topic method.