Research On Content Based Image Retrieval Techniques In P2P Environment

With the development of computer and communication technology, digital image retrieval is widely used in varities of fields. Traditional information management and retrieval methods are not suitable to large image databases any more. In order to manage and retrieve large amount of images, the CBIR (Content-Based Image Retrieval) has emerged as a hot research topic in image field.CBIR is executed on basis of images' content, such as color, texture and shape, which are extracted by computers itself. The system retrieves the feature vectors consisting of the above information from large image databases to meet the users' requirement.This thesis introduces the system architecture and some key techniques together with CBIR. Image clustering strategy and routering strategy are deeply researched in this dissertation.The diamention of feature vectors of images can reach up to tens or even hundreds easily, which makes the traditional index strategies work awkwardly. So, how to index and search effectively becomes another research topic. Most of traditional image retrieval systems, including those famous ones, such as QBIC, Visual SEEK, etc. are based on centered database. How do make the most of digital images distributed in the internet is the second issue we considered in this paper. We try to take the advantage of precise resource location of DHT (Distributed Hash Map), as well as the unsupervised, self-organizing and steady TS-SOM (Tree Structured Self-Organzing Maps) to attack the above two problems. Primary experiment results show it works well.How to apply DHT into CBIR to do semantic search is an interesting topic, we did the following exploration to try to attack this problem: Yingwu Zhu uses fixed m and n to classify files in his LSH-based semantic search system, satisfactory results can hardly be reached when similarities varies wildly. As we know, neighborhood nodes have higher similarity in CAN, so we take advantage of this feature, and cache strategy is introduced into this sytem to calculate file similarity and get the optimized m and n. We prove it is promising, theoretically.