Research On Local Region Detection Methods And Its Application In The Filed On Image Retrieval

The key technology of CBIR(content-based image retrieval) is how to extract and describe features from an image. Because of changes in lighting, scales, object pose, camera view, the complex environment and other factors, at present, the image representation based on global features can’t meet the image retrieval demands. To make up the shortcomings of the global features, the researchers have proposed the image representation methods based on affine invariant region. However, the present region detection methods will detect many background regions when the background of the image is complex. If we don’t remove the background regions, it will not only affect the results of image retrieval but also increase the processing time. In addition, these methods are not make use of the human visual characteristics, so the region don’t have the real salient. Therefore, how to effectively detect the regions in accordance with human visual characteristics is a very important problem.This paper proposed a local region detection method which is incorporate image structure with information entropy based on the exiting region detection method. On the side, presented a salient region detection method based on the characteristics of human vision and proposed a image retrieval system based on image local regions. Our main work includes the following five aspects:1. Combined the image structural features and information entropy to detect the image regions, which made the detected image regions with steady and significant.2. In order to accord with the human visual characteristics, this paper used the visual priori knowledge and biomimetic information theory.3. Learned from the current popular SIFT descriptors to describe the image regions, which made the region descriptors have better robustness.4. To improve the efficiency of the image retrieval, used the hierarchical K-means clustering algorithm and vocabulary tree index structure.5. In similarity measure, this paper adopted the voting score to achieve the image retrieval based on image regions.