Fusion of color and statistic features for enhancing content-based image retrieval systems 1

Content-based image retrieval is one of most debated topics in computer vision research, and has received a great deal of interest recently. It aims to retrieve similar images from a huge unlabelled image database. In this work we propose a method that reduces the error rate and retrieves relevant images early in the process, with the ability to work on both color and grayscale images. The proposed method scans an image using 8x8 overlapping blocks, extracting a set of probability density functions of the most discriminative statistical features. Our experiments, conducted on several image databases, show the robustness of the proposed method, outperforming some of the most popular methods described in the literature.