A novel ranking algorithm based on manifold learning for CBIR system

At present, most of image retrieval applications use Principal Component Analysis (PCA) algorithm to reduce the low-level features of images and rank the similarity of images based on graph structure to enhance the retrieval accuracy with relevance feedback technique. However, there are two issues to consider in traditional image retrieval methods: (1) the feature space of images is probably highly non-linear, in this case, PCA always fails to uncover the intrinsic structure so that the performance of dimension reduction is unsatisfactory; (2) ranking algorithms based on manifold learning most likely ignore the global structure of image feature space. To address the issues above, this paper utilizes Linear Local Tangent Space Alignment (LLTSA) algorithm to uncover the non-linear structure of images feature space. At the ranking stage, we take advantage of the semi-supervised idea to sort the similarity of images. Such a strategy makes up the shortcomings of manifold ranking. Experiments on a large collection of images have shown the effectiveness of our proposed algorithm.