Regression Models for Relevance Feedback and Feature Selection in Content-based Image Retrieval

One of the main reasons why content-based image retrieval (CBIR) is such a challenging problem has to do with the fact that there is no canonical way to capture the visual content encapsulated in an image. As a consequence, most CBIR search-engines keep the human in the loop by regularly requesting his feedback to expedite the search-action. More precisely, the feedback is harnessed to estimate for every image in the database, the likelihood of its relevance to the user, whereupon the most promising candidates are displayed for further inspection and feedback. This procedure is then iterated as often as necessary to locate the target image. To model the fuzzy state of knowledge about the user’s preferences, the search engine assigns to every image Ij (j = 1; : : : ; N) in the database a relevance probability p(Ij) that reflects the current estimate of relevance. Hence, p(I) 1 means that in its current estimate, the image I is considered to be highly relevant, whereas p(I) 0 expresses the opposite. Initially every image in the database is considered equally likely. However, as more information about the user preferences becomes available (through relevance feedback), the probability measure will start concentrating on regions in the database that seem promising, and images in these regions are more likely to be sampled for display.