A dynamic sub-vector weighting scheme for image retrieval with relevance feedback
In image retrieval with relevance feedback, feature components obtained from low-level descriptors are often 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 the labelled images is often small and an optimal weighting cannot be achieved in practice because of the singularity of the covariance matrix needed for weighting. To solve this problem, a dynamic sub-vector weighting scheme is proposed in this paper. In this scheme, a multidimensional feature vector is partitioned into multiple sub-vectors whose dimensions are dynamically adjusted to match the small number of available labelled images. Because of the lower dimensionality of the sub-vectors, the optimal weighting can be performed based on these sub-vectors, respectively. The multiple similarity scores obtained from sub-vector comparisons are then combined, as the final score, to rank the database images and determine the retrieval result. Experimental results demonstrated better performance of the proposed weighting scheme than some existing weighting schemes.