A SUB-VECTOR WEIGHTING SCHEME for IMAGE RETRIEVAL with RELEVANCE FEEDBACK

In image retrieval with relevance feedback, feature components are weighted to reflect the high-level concepts, and a user's subjective perception, embodied in the images labelled by the user in the feedback. However, the number of labelled images is often small and the covariance matrix needed for weighting will be singular. For this reason, the commonly used methods discard the mutual correlation among the feature components completely and use a diagonal covariance matrix. In this paper, a sub-vector weighting scheme is proposed. This scheme partitions a multi-dimensional visual feature vector into multiple low-dimensional sub-vectors. The singularity of the covariance matrix for each sub-vector can be avoided due to the lower dimensionality of the sub-vectors. Thus, the mutual correlation in each sub-vector can be retained for weighting and an optimally weighted similarity metric can be applied on each sub-vector. The similarity scores obtained from different sub-vectors are combined, as the final score, to rank the database images. Experimental results demonstrated that the proposed weighting scheme can significantly improve the efficacy of image retrieval with relevance feedback. © 2002 World Scientific Publishing Company.