The Study Of Image Retrieval System Based On Relevance Feedback And Particle Swarm Optimization

With the rapid development of electronic devices and social networks, millions of images are shared in the Internet every day. To retrieve desired images from a large-scale image database efficiently, content-based image retrieval(CBIR) technology has always been a hot topic of many scholars.CBIR has overcome the disadvantages of traditional text-based image retrieval, which is subjective and requires a heavy workload. However, due to the semantic gap problem between low-level features and high-level semantic meanings, the retrieval results are difficult to satisfy all users. Therefore, Relevance feedback(RF) is introduced into the field of CBIR to solve these problems; nevertheless, the current RF mechanism requires experienced users to label a large number of images, which brings heavy burden to them. Moreover, the process of RF is based on certain algorithm, which makes the retrieval process fall into local optimum easily.In order to solve these problems, this paper mainly studies relevance feedback and particle swarm optimization(PSO) in the field of image retrieval. The work in this thesis mainly include:(1) The diversity of queries is limited in current PSO-based image retrieval approaches. In this thesis, the multi populations of PSO(MPSO) are introduced into RF and the approach of image retrieval based on RF and MPSO is proposed. Firstly, the points are found, which have high diversity in relevant image set, then multi populations of PSO are initialized according to these points. The proposed approach can improve the diversity of PSO particles, thus can improve the diversity of the queries optimized by MPSO.(2) Previous research has founded that there are still a lot of redundant information in the queries optimized by MPSO. In order to improve the diversity of the queries and reduce the redundancy, with some filtrations, the new queries are got from the particles which are selected from each population. We also study how to sort the multiple image sequences retrieved by the queries, and proposes a novel sorting method which has combined the fitness of queries and the similarity of images in each image sequences.(3) We analyzes the reason why there is much redundant information and the burdens on users in image retrieval based on traditional RF, and proposes a novel RF framework combining MPSO. After three filtrations, the image set to be feedback is selected form the similar image sequences retrieval by every query. The novel RF mechanism, which calculates the retrieval results and the image set to be feedback separately, not only improves the effectiveness of feedback information, but also reduces the burden on users greatly.