Research On Image Retrieval Technology Based On Domain Clustering And Feature Fusion

Region-Based Image Retrieval (RBIR) technology is an important research direction of content-based image retrieval (CBIR), in which image segmentation technology is used to divide image into several region and the region features are utilized to represent, index and retrieve the images. To some extent, RBIR realizes the retrieval of the images in the target (or object) level, which dramatically reduces the semantic gap between the low-level image feature and the high-level semantic concept. Currently, Image search is an important application of RBIR.Image segmentation technology is the focus and the difficult in RBIR. Based on the analysis of the drawbacks of the traditional segmentation algorithms, we attempt to combine multiple algorithms together to complement each other. The combination of watershed region growth and the minimum spanning tree achieves the most consolidation of segmentation. However, watershed and C means algorithm has better interoperability, high efficiency and more suitable for the characteristics of visual perception of human by considering it is used in the image retrieval application. The combined algorithm has good consolidation segmentation results and high efficiency.As the retrieval is based on region in RBIR, the feature extraction and the matching strategy is different from the general image retrieval system. Multiple features, including choice color, texture and region size, are extracted and adopted to index the images. The image information and the region features are indexed separately for better modularity. The EMD algorithm, which is originally for solving transport problem, has been used to calculate similarity between region-based images for its fast match speed and strong scalability.A region-based image retrieval platform has been developed based on the above key technolgies. It has abundant functions, such as establishing image index files for image data base, providing image retrieval for the users, deriving target image to a specific directory. Based on the experiments on the platform, which includes segmentation, feature extraction and matching algorithm,the results show the effectiveness of the region-based image retrieval algorithms.In addition,the comparasons between global feature based and region feature based retrieval, a single segmentation algorithm and the combined segmentation algorithm, multiple features and single feature for the retrieval,have shown that the retrieval precision and the efficiency of the region-based image retrieval algorithm are remarkedly improved.