Unbalanced region matching based on two-level description for image retrieval

Research on integrating spatial information into content-based image retrieval (CBIR) is aimed at solving the problem caused by global feature based algorithm. Most systems derive the spatial information from image segmentation. However, the description of images based on one-level segmentation (OLD) and the inevitable inaccuracy of segmentation results seriously limit the performance. A two-level description (TLD) describes images by a rough description and a detailed description to avoid improper spatial constraint caused by OLD is proposed. Similarity measurement based on unbalanced region matching (URM) is introduced to take advantage of TLD to reduce the influence of segmentation. A novel spatial descriptor integrating shape, size, and density as well as position and spatial layout information together is also proposed. The performance of the integrated system is illustrated by experimental results with 1000 query images randomly selected from a database of 10,000 general-purpose images.