Abstract: Existing algorithms based on structural descriptor are not accurate enough to discriminate the image features, because they lack the correlation description of continuous pixels or structural elements. To address this problem, this paper presents a novel weighted quantization method, which can adaptively integrate images object features and background features into one image histogram. The proposed method includes the new structure elements definition, adaptive vector fusion model and connected granule concept. Firstly, based on the visual selection characteristics, nine kinds of new structure elements are defined. The connected granule’ attributes are given and the hierarchical statistical model is constructed. Secondly, the corresponding mapping sub-graphs are generated by color transformation and structure elements matching. Meanwhile, the feature vectors of statistical structure elements and connectivity are extracted. Finally, a set of feature vectors are obtained by utilizing adaptive vector fusion model for image retrieval. Extensive experiments on three Corel-datasets demonstrate that the novel algorithm performs better than several state-of-the-art image feature representation methods. The proposed method not only captures the global features of images, but also reflects the local details of images.