The Study Of Image Retrieval Based On Indexing For Feature Of ROI

Content-based image retrieval technique has become a hot research area, in which much outstanding achievement has been made and on which the research developed rapidly.The early research in the area mainly focused on image retrieval based on feature similarity.Many typical content-based image retrieval system performed description and similarity measurement based on color and texture feature of the whole image to realize related image retrieval ,however,the performance of these retrieval couldn't satisfy user's demand.In fact,we human beings don't measure the similarity between two images only in terms of the feature .The original idea on content-based image retrieval is that it could be realized in terms of understanding of image and knowledge on image,which represents semanteme.To provide user with natural and brief retrieval modes in order to improve retrieval accuracy more,image retrieval must be realized based on semanteme which could be obtained by feature of image.So,there are problems :one is that the description modes of semanteme of image must be provided,another is that the mapping method from feature to semanteme must be aquired.Now,there is not conjunction between visual feature and semanteme essentially and the'semanteme gap'exists still.How to extract semanteme from image is the most challenging task in content-based image retrieval.In my opinion,the first level for the outstanding of image is the knowlwdge on Interest,typical,semantic region in image,which is ROI(Region Of Interest) ,then,the layout among these ROIs .So the extraction and analysis of typical semantic region is the base of the semantic analysis for the whole image. The extraction of typical semantic region is segmentation of image essentially,however,now,there is no typical,general,effective method for segmenting image automaticly.In the dissertation,it is proposed that typical semantic object region is extracted from image by operator's intervention and as little interaction between operator and computer as possible ,which are called as image primitive and are the base for content-based image retrieval.In the dissertation,first of all,image primitives are extracted from images of complex content through recursive region-growing by interaction between operator and computer based on color and texture combined feature,and these primitives can be combined to obtain bigger primitives according to semantic correlation and spatial near neighbour.Then,the size of image premitives is standardized to extract feature, which is used to learn ,cluster and recogniztion gradually and supervisedly images by RSOM(Recursive Self-Organizing Map) and to form a'knowledge-base'on image primitive .Finally,image primitives or their content as'image keywords'being associated with original images,like text indexing,image indexing by image primitives or their content,in the base of clustering and recognization of image primitives,a kind of intelligent,coincident in visual perceiving image understanding and retrieval is realized.