Research On Attributed Graph And Clustering Tree Based Large Datasets Image Retrieval

Image retrieval is an active research domain in computer vision. Although some successful progress has been achieved recently, it is still a challenging problem because of the complexity, especially for large image datasets. This thesis summarizes current methods of image retrieval and analyses the deficiencies of these methods.Technologies of image representation, image indexing, and retrieval policy have been thoroughly studied and the corresponding algorithms are implemented.In Chapter 1 the up-to-date image retrieval technologies are introduced, and some challenging problems of current technologies are proposed. Then research scheme and the work of this thesis is briefed.In Chapter 2 all fundamental theories used in this thesis, including local invariant feature extraction, construction of attributed graph using local invariant feature, and RSOM tree, which is used to index images, is introduced. As follows, similarity propagation mechanism, and Class Specific Hyper Graph (CSHG) model, and the construction method of CSHG are briefed.In Chapter 3 the detection algorithm of K nearest neighbors (KNN) of images based on RSOM tree is described, including: image indexing, PKNNG detecting, and KNNG detecting. As for high-dimension feature, RSOM forest based segmenting feature modeling method is proposed. RSOM Forest training and detecting algorithms and the RSOM Forest KNN detection algorithm are introduced.In Chapter 4 image index construction methods in this thesis is introduced, and similarity propagation based image retrieval is realised. Based on public academic testing datasets, some experimental results have proved the efficiency of our methods.Finally, the work is summarized and prospect of image retrieving in large datasets is pointed out in chapter 5.