A relevance feedback-based learner for image retrieval using SIFT descriptors

Abstract

This paper presents a relevance feedback (RF) algorithm for scale invariant features extracting form Caltech image database. The RF is a powerful technique to bridging the gap between high-level concepts and low-level features in image retrieval systems. This paper attempts to enhance the performance of RF by exploiting unlabelled images in the database. Each scale invariant feature transform (SIFT) keypoint is considered as the different feature (colour histogram, colour moments, wavelet transform, Gabor transform, etc.). The user labels several images accordingly to whether they are positive (relevant) or negative (irrelevant) examples to a query. Therefore, the relevance feedback process changes labelled and unlabelled images. The query feature and weights are updated as per the keypoint features of positively and negatively labelled images. Furthermore, a new similarity measurement for matching of database images with query is proposed in this paper. The SIFT feature is local feature of an image, which is invariant to image scaling, transformation, rotation and partially invariant to illumination changes, angle of view changes, and noise. The retrieval results show that proposed RF method using SIFT descriptors achieves good effectiveness in content-based image retrieval.