An Optimal Codebook for Content-Based Image Retrieval in JPEG Compressed Domain

Images in JPEG compressed domain have a widespread use and computationally efficient retrieval of those images is of prime concern. In this paper, a content-based image retrieval system in JPEG compressed domain is presented, which generates an optimal codebook and extract features that only require partial decoding of images. The codebook is generated by selecting an optimum number of training images and feature vector length using an optimization cost based on precision and recall. The cost gives the difference in average precision and average recall, which the retrieval system can tolerate, while reducing the codebook size and feature vector length. The efficiency in retrieval performance is achieved by generating an optimal codebook. Experimental results have shown better performance for the proposed retrieval system in terms of precision, recall and number of operations, when compared with state-of-the-art image retrieval systems in JPEG compressed domain. The proposed retrieval system also shows robustness against compressed query images by using different quantization parameters.