Efficient image retrieval in DCT domain by hypothesis testing

We consider a hypothesis testing approach to content-based image retrieval (CBIR) using discrete cosine transform (DCT) coefficients restored by partially decoding JPEG images. In order to further decorrelate DC coefficients from an image, a 2 × 2 DCT is performed on the sub-image constructed from all the DC coefficients. Assume that each DCT coefficient sequence is emitted from a memoryless source, and all these sources are independent of each other. For each target image we form a hypothesis that its DCT coefficient sequences are emitted from the same sources as the corresponding sequences in the query image. Testing these hypotheses by measuring the log-likelihoods leads to a simple yet efficient scheme that ranks each target image according to the Kullback-Leibler (KL) divergence between the empirical distribution of the DCT coefficient sequences in the query image and that in the target image. Experiments on two image datasets show that our approach achieves consistently better retrieval results than related methods in the literature.