Machine Learning For Content Based Image Retrieval

With the popularization of image capture devices and the reduction of storage cost,lots of unorganized digital images pile up both in PCs and on Internet. The research one?ciently organizing, retrieving, and browsing them has become crucial.This thesis aims at reducing user interaction in the application of content-basedimage retrieval, boosting the performance of the retrieval algorithm, and improvingoverall user experience, by solving three related machine learning problems in CBIR.Firstly, when the relevance feedback is label of the whole image, the user is al-lowed to label multiple images in a batch. The problem of selecting the optimal batchis solved using sampling based active learning. Compared with traditional greedy opti-mization based methods, user interaction required to achieve comparable performanceis reduced.Secondly, a new relevance feedback mechanism of drawing strokes inside imagesis proposed, allowing the user to specify interested regions more intuitively. Betterresults are achieved with more user specified information. An unified framework ofcombining the new mechanism with the traditional one is also proposed.Thirdly, to further improve the performance of algorithms based solely on visualfeatures, this thesis proposes to first automatically label the images with textual tags,then use them to retrieve. These tags are more close to the semantic meanings of theimages, thus overall performance of the retrieval algorithm can be improved.Moreover, this thesis introduces the image retrieval system we designed and im-plemented during the research. As a general platform, the system supports typicalfunctionalities of image search engines including crawling, indexing, feature extrac-tion, retrieving, etc. The algorithms in all steps are modulated, which make it veryeasy to change the algorithm of any step.