DCT-based content descriptors for JPEG compressed image indexing and retrieval

In this paper, we exploit the JPEG compression technical properties to propose simple content descriptors for content-based image indexing and retrieval, which is particularly suitable for indexing JPEG compressed images. Given an image divided in N blocks of 8x8 pixels, we construct an indexing key by overlapping the N blocks into one combinational block and each block acting as one single plane inside the combinational block. This is prompted by following, the spirit of bit-plane encoding of images, where each pixel is divided into a bit sequence and every bit of all pixels constructs a bit plane. As a result, the indexing key will have 64 elements and each element corresponds to one location along the zig-zag route of the JPEG compression standard. Specific construction of each element inside the indexing key can have three alternatives and thus three content descriptors can be constructed, which include energy-based content descriptor, binary texture based descriptor, and orientation based descriptor. Extensive experiments are carried out and the results support that the proposed texture descriptor produces competitive performances in comparison with existing techniques. Yet the proposed texture descriptor illustrates a number of advantages and features, which can be briefly summarised as: (i) low complexity and low computing cost; (ii) high processing speed, particularly suitable for texture classification or clustering inside large image databases; (iii) easy to implement inside JPEG compressed domain, and thus providing additional advantages that compressed images can be directly retrieved without full decompression.