Image Retrieval Based on Hierarchical Locally Constrained Diffusion Process

To address the problem of high matrix computational cost in traditional diffusion process methods for large scale image retrieval, we propose a novel image retrieval algorithm based on hierarchical locally constrained diffusion process (HLCDP) by combining algebraic multigrid and diffusion ranking on manifold. In HLCDP, all retrieved images of the image database are built into a bottom-to-top hierarchical structure by selecting the representative images. Then the similarity among images on the top layer is diffused by using locally constrained diffusion process, and the affinities (i.e. ranking scores) between query images and top-layer representative images are interpolated to all images on the bottom layer to obtain the affinities between the query images and all of the images in database. Our method is evaluated on image retrieval tasks by comparing with locally constrained diffusion process (LCDP), self-smoothing operator (SSO) and self-diffusion (SD). The experimental results on MPEG7 data set and ImageCLEFmed2005 data set demonstrate that the proposed method improves the retrieval performance and reduces the retrieval time consumption.