Category-adaptive color image retrieval based on Lloyd-clustered Gauss mixtures

Traditional image retrieval systems consist of two working blocks for a query image: Feature extraction and similarity matching. This structure, however, does not provide satisfactory retrieval performances for different image databases. Therefore, image retrieval systems based on users' relevance feedback have been developed. Since this structure assumes the relevance feedback, it cannot be applied to fully automated systems. It can also annoy the users who may be reluctant to repeat the querying process. Therefore, we propose a better automated image retrieval system, which adds a classifier to the traditional system for giving feed-forward information to maximize the average precision. It requires a pre-analysis of the given image database.; We find that every category of images has its own mode (a combination of color space, quantization method, and similarity measure) leading to a better average retrieval performance than other modes. That is, there is no single mode to maximize the average retrieval performance for every category of images. Given a database, our proposed system will improve the overall precision by selecting the best mode based on known statistics (average precision vs. recall for each category). The Lloyd-clustered Gauss mixtures are used in the classifier to provide the feed-forward category information and in the quantization of color images for histogram generation.; The proposed system improves the retrieval precision by about 2.8 percent over our previous best single mode retrieval system. The result is better than that of the famous color indexing of Swain and Ballard (91') by 9 percent for the generic color images used. The proposed system can improve the retrieval precision by up to 5.5 percent over the best single mode retrieval under our experimental setup, given a perfect classifier.