Feature Extraction in Compressed Domain for Content Based Image Retrieval

Content based image retrieval (CBIR) systems in pixel domain use low-level features such as color, texture, and shape for image queries. Extracting these features and comparing it with the database images is time consuming. Also majority of the images stored in the systems are in JPEG compressed format. So, it would be highly desirable if we can extract the features directly in compressed domain, and use these features for retrieval purposes. In this paper, we present an algorithm to extract the features directly in the compressed and uncompressed (YUV) domain. Novel usage of features like skewness and kurtosis along with the standard set of statistical features helps in discriminating the images more accurately. In addition, the system has the property of robustness to rotation, scaling, translation, and illumination correction. Experiments on both object and facial database demonstrate that the proposed method increases the retrieval speed by 10% and reduces the required memory to store the features by 25 %. It also improves the retrieval accuracy significantly when compared with the conventional algorithms.