The Research And Application Of Structural Image Representation

Managing, classifying, and retrieving the multimedia data are the important task in the field of computer vision, which have broad application values and urgent prac-tical needs. Effective image feature representation plays the fundamental role in im-plementing the above computer vision tasks. In the past, human obtains the subjective, semantic meanings of the image mainly according to their experience and knowledge. Consequently, the visual features extracted by computer fail to describe the high-level semantic meaning accurately, which raised the problem of " semantic gap " between the low level feature and the high-level semantic meaning. The research of image rep-resentation is an effective method to solve the "semantic gap" problem, but existing methods still suffer from the problems of weak representation ability and discrimina-tion ability. To address these problems, we need to learn the structural image feature representation. In this thesis, we proposed some methods to obtain the structural image representation. Moreover, these methods have achieved the good results in the image retrieval and classification tasks. The main work of this thesis is as follows.Firstly, we present an regularized method of the enhanced relational matrix to im-prove the traditional semi-supervised manifold learning algorithm. According to the relational matrix of manifold learning based on graph embedding, we find the neigh-borhood relationship of the data and enhance the neighbor’s neighbor. Then we reg-ularized the relational matrix using the transition probability matrix and propagate the neighborhood information among the entire data sufficiently. We reduce the dimension of the image feature by the projection vector. Finally, we obtain the structural image representations. With the structural image representations, we implement the experi-ment of feedback image retrieval. Experimental results show the efficiency of the new algorithm.Secondly, According to the problem of the codebook generated by the BoW mod-el, we present an structurally preserved incremental neural network method. This method generates the codebook graph and provides richer information for feature en-coding. The incremental neural learning can learn appropriate visual words set adap-tively via online approach. With the relationship of the visual words, we design an subgraph-based coding method to encode the local features and obtain the structural image feature representation. In the experiment of image semantic classification, the results show that the efficiency and accuracy of image classification are improved by our method.Finally, we proposed an adaptive locality-constrained linear coding method based on the enhanced structural codebook and developed a geometric smooth pooling scheme to generate the structural image representation. In the coding phrase, the codebook is initialized by the approximated K-means with cluster closures, greatly decreasing the computational complexity. The codebook is updated by exploiting the local geomet-ric structure of the code atoms. It encodes the image feature by adaptively locality-constrained linear coding method. In the pooling process, we also develop a geometric smooth pooling strategy based on the spatial pyramid framework aiming to make full use of the rich information about spatial distribution of the extracted features. Finally, we obtain the structural image representation. In the experiment of image semantic classification on the benchmark dataset, we get the high classification accuracy and verify the effectiveness of the proposed algorithm.