Relevance feedback techniques for color-based image retrieval

Abstract

Color has been widely used in content-based image retrieval systems. The problem with using color is that its representation is low level and hence its retrieval effectiveness is limited. This paper investigates techniques for improving the effectiveness of image retrieval based on colors. It examines the choice of suitable color space and color resolution. It describes two techniques for image retrieval with relevance feedback (RF). The first uses machine learning algorithms to extract significant color intervals and build the decision tree from the relevant image set to support effective RF. The second employs color coherent vector (CCV), in which the pseudo object information encoded in CCV is used for RF. Both techniques have been tested on a large image database containing over 12000 images. Tests were also performed to evaluate the effectiveness of retrieval at different color resolutions. The results demonstrate that our RF techniques are effective and a medium color resolution of 176 colors performs the best.