Study On Key Techniques Of Region-based Semantics And Features Image Retrieval

Several key techniques corresponding to m a in six steps in region-based im a g e retr ieva l are introduced in detail. Also com b ing with active resear ch in this area, a n approach of building region and im age se m a ntic networks and applying relevance feedback for im proving on retrieval ef ficiency of based on the low-level features is presented. These steps are feature extracti on, im age segm entation, region descriptions, im age m a tc hing, sem a ntic networks, and re levance feedback. An i m age is firstly × partition e d into blocks with 4 4 pixels, extracting a featur e vector for each block. The feature vector is com p osed of eig h t feat ures; those are 3 colo r features, 3 texture features and 2 position features. Such integra ting feature vector is used for building K dim e nsion gaussian m odel, whose param e ters are estim ated by an expectation-m a xi m i zation (EM) algorithm , and then the resulting block-cluster m e mberships provide a segm entation of th e im age. After segm ented, a m e thod of param e ter - trimm e d average for describing re gion is proposed, of which the param e ter is decided by area and position of region dire ctly . The sim ilarity m easure between two im ages is defined by integrating properties of all regions in the im age. Com p ared with retrieval based on individual regions, the appr oach reduces the influence of inaccurate segm entatio n and prov ides a v e ry intuitive qu antif ica tion. In or der to im prove the index precision, the keywords of classification on im age database is used for for m ing a three-level sem a ntics network, and then a unified fra m ework for sem a ntics and feature based relevance feedback in region-based im age retrieval is described, which is experim e nted by irrelatively adjusting the keywords and the weights in the distance m easuring.