Content-Based Image Structure Representation And Classification

The thesis focuses on content-based image structure representation and classification. As a widely used technology, CBIR has been given more and more attention, and many achievements have been developed and applied in real world. However, there still exist key issues in the research domain of content based image retrieval and classification. Recently, a fewer of researchers took into account that Recursive Neural Network (RNN) can be employed to represent the structural patterns and implement the classification of data structure by using BPTS algorithm. The aim of the thesis is to study the key problems of feature extraction, pattern representation and classification for complicated patterns.In feature extraction for image retrieval, we systematically review and summarize the proposed approaches which cover the research area for traditional feature extraction, for example, color, texture and shape etc. Experimental results have shown that HSV color system is more adaptive to people perception compared with RGB color system and a novel feature extraction method is proposed, which use a new partition strategy to divide the natural scene image into five blocks and local statistic features are extracted.Moreover, the framework and principle of adaptive processing of data structure are discussed in detail. The representation of data structure and learning algorithm for DAGs by back propagation through structure are also represented in the thesis.Conventionally, the natural scene image can be represented by four-branch trees according to hierarchical partition. In the thesis, the image is represented by multi-branch directed trees manually or binary trees automatically based on the segmented result using Berkeley segmentation algorithm. The tree structures and the statistic features related with the structure, including RGB components and HSV bins are used in learning and classification. Experimental results have shown that the recursive network based structure model has significant ability to represent complicated patterns and the multi-branch tree has more semantic meanings to achieve higher classification rates.Finally, the classification performance is analyzed corresponding to the net structure and the number of hidden neurons.