Research Of Two-stage Attention Driven Image Retrieval Approach

This dissertation discusses some key problems on content-based image re-trieval technic focusing on shortcomings of current systems in e?ciency and pre-cision. An attention driven two stage image retrieval scheme is presented, em-phasizing on feature extraction, salient objects segmentation, and region basedsimilarity measure of the system. Based on the discusses above, an integrity metasystem is described, and its e?ectiveness is proved in the experiments.Problems studied include: First, retrieval is classi?cation in essence exceptthat one or more patterns are gained in a retrieval task compared to recognition.Based on this concept, we conclude a set of pattern recognition technics suitablefor retrieval task, and pre-classify images into categories before ?ner retrieval. Itsolves the e?ciency problem reasonably. Second, most current systems employglobal feature for retrieval, and thus, background which possesses most area of animage determines the result of retrieval. Users often confused with these results,and it leads the relevance feedback to the opposite direction. An attention drivenimage segmentation approach is proposed in this dissertation based on regionbased retrieval technics. It enables region based retrieval being guided under abottom-up computational model of human visual attention. As a consequence,precision problem is settled. At last, region based similarity measure technics arediscussed.Main innovations of the dissertation are:? Splitting the whole process of retrieval into two stages in logic. E?ciency isgreatly enhanced by classifying image in the database before hand.? Methods that extract region of interest blindly are abandoned. Making theprocess of image segmentation accord with the selection of visual attention.Precision problems are solved by this conception.