A new approach for relevance feedback through positive and negative samples

Relevance feedback has recently emerged as a solution to the problem of providing an effective response to a similarity query in an images retrieval system based on low-level information such as color, texture and shape features. This work describes an approach for learning an optimal similarity metric based on the analysis of relevant and non-relevant information given by the user during the feedback process. A positive and a negative space are determined as an approximation of the examples given by the user. The relevant region is represented by a KL subspace of positive examples and is iteratively updated at each feedback iteration. The nonrelevant region is modeled by a MKL space, which better characterizes the variety of negative examples, which very likely could belong to more than one class. The search process is, then, formulated as a classification problem, based on the calculation of the minimal distance to the relevant or non-relevant region.