An Image Retrieval Method Using DCT Features

In this paper a new image representation for compressed domain imageretrieval and an image retrieval system are presented. To representimages compactly and hierarchically, multiple features such as colorand texture features directly extracted from DCT coefficients arestructurally organized using vector quantization. To train the codebook,a new Minimum Description Length vector quantization algorithm is usedand it automatically decides the number of code words. To compare twoimages using the proposed representation, a new efficient similaritymeasure is designed. The new method is applied to an image database with1,005 pictures. The results demonstrate that the method is better thantwo typical histogram methods and two DCT-based image retrieval methods.