Based On Local Sensitivity For CBIR Research

With the rapid development of the Internet, the number of images on the Internet also hasa rapid growth. Effective image retrieval from media databases have become a researchhotspot. There are two major techniques for image retrieval: keyword-based image retrievaland content based image content retrieval (CBIR). The former technique is relatively mature,but also has many shortcomings. For example, keyword-based methods need manuallytagging of keywords to images which is impractical for web based large scale image retrievalproblems. The CBIR was proposed to retrieve images based on image content directly torelieve the needs of image tagging.In this thesis, the basic knowledge about the CBIR and the Localized GeneralizationError Model (L-GEM) are introduced. The sensitivity measure in the L-GEM has been widelyused in pattern recognition problems for classifier optimization. One of the majorcontributions of this thesis is the introduction of the sensitivity measure to CBIR problems.In image similarity learning, a tiny modification of image may cause disturbance in thefeature space. Hence, we enhance the OASIS method published in the JMLR by adding thesensitivity measure to its objective function for a robust similarity metric learning.Experimental results show that the proposed method achieves better results.On the other hand, the sensitivity measure is also applied to enhance the querymovement method for CBIR problems with relevance feedback. The sensitivity measure isused as the indicator to update centers of both the relevant class and the irrelevant class.Experimental results show that the proposed method yields a higher accuracy in comparisonto methods without using the sensitivity.Finally, we summarize methods in this paper, and then put forward some prospects offuture research works.