A fast approach for identifying similar features in retrieval of JPEG and JPEG2000 images

As digital images are often in compressed forms, image retrieval involves full decoding of images prior to featureextraction. The decoding process can be computation-expensive so feature extraction in compressed domain is desired. In this work, wavelet-based features are extracted as unified features for retrieval of JPEG and JPEG2000 images. A fast algorithm is proposed to approximately transform a JPEG image in the block-based discrete cosine transform (BDCT) domain to wavelet domain so that wavelet-based features can be extracted directly from JPEG images. Our proposed algorithm consists of a multiresolution reordering and a filter bank structure. The former is used to provide a rough approximation of wavelet subbands from BDCT coefficients in bandpass subbands in fine scales while the latter is used to provide an accurate approximation in bandpass subbands in coarse scales. Our theoretical analysis shows that the proposed algorithm can reduce the complexity by at least 79% when comparing with the straight forward approach that uses an inverse BDCT followed by wavelet transform. Besides the reduction in computational complexity, the experimental results demonstrate that our proposed conversion approach has higher retrieval performance than the pure multiresolution reordering approach.