Boosting retrieval efficiency with image replacement based relevance feedback

Relevance feedback has been employed in Content Based Image Retrieval systems to bridge the semantic gap between the low level features and high level semantics of the image. This paper proposes a short term learning relevance feedback algorithm that utilizes the statistical features of the feedback images for determining the relevance of the candidate image in the next iteration and for achieving improved precision. The similarity of the candidate image with the feedback image set is determined by computing the cumulative sum of the displacements of the feedback image centroid caused by replacing each element in the feedback image set with the candidate image in the database. Experimental results show that using the proposed image replacement algorithm, improved precision of 8% can be achieved even with a single image given as feedback by the user. Also it is seen that optimum number of feedback images needed for obtaining improved performance is 2-10.