An improved method of image retrieval with relevance feedback using random walks

Despite the efforts to reduce the so-called semantic gap between the user´s perception of image similarity and the feature-based representation of images, the interaction with the user remains fundamental to improve performances of content-based image retrieval (CBIR) systems. It is fact that the effectiveness of a CBIR system strongly depends on the set of visual features and the `metric´ used to model the user´s perception of image similarity. Random walks for CBIR with relevance feedback (RF) introduced by Samuel Rota Bulo^[1] is a relatively complete retrieval system. However, it does not take into account the importance of each dimension characteristic and image spatial relations when using the random walk. In this paper, the problem is solved by two strategies: firstly, feature weighting using Laplacian Score, which is one of the classical approaches to implement RF, and secondly, K-NN connecting to Random walk, which uses the K-NN of relative and non-relevant images during RF to improve the efficiency of random walk. The performances of the proposed method have been compared with different recent relevance feedback mechanisms. Reported results have shown the superiority of the proposed method.