Abstract: Most existing content-based image retrieval systems using low-level features that could not describe high-level semantics thoroughly and accurately. In this paper, a novel system for content-based image retrieval is designed and created, which combines image semantics based on a multi-level model for image description. In this image description model, image contents could be analyzed and represented through different levels and the transition from low-level features to high-level semantics is thus achieved. Corresponding querying mechanism and feedback are also proposed based on this image model. Aiming at object semantics in image, this querying mechanism is much closer to human beings understanding of image contents so that it provides a convenient and effective querying procedure. The feedback used in the system is a self-adaptive relevance feedback based on object descriptions, it permits to propose different querying schemes according to the different demands raised by various users, and thus optimal results could be refined.