An image content-driven (CDP) preprocessor is proposed to activate the right MPEG-7 description tools for the recognized feature contents in one image. It determines automatically whether there are certain feature contents, such as color, texture or shape features, in one image and then performs processing to generate the corresponding descriptors. The CDP's most distinguished characteristic is that there are no redundant computations from the image content categorization down to the descriptor generation. Experiments show that the proposed CDP framework effectively categorizes images with accuracy up to 99%. We also proposed a practical content-based image retrieval (CBIR) system which integrate the CDP framework with a user-friendly MPEG-7 testbed. Simulations of CDP-based CBIR demonstrate that the CDP helps much in improving the subjective retrieval performance. This CBIR framework provide excellent _exibility such that it could be easily adapted to meet speci_c application requirements.