Image Retrieval Based On Multiple Random Walks

With the rapid development of digital camera and network technology, the types and the information of image data are both becoming richer. In order to find the necessary information from these rich image data quickly and efficiently, image processing technology need to grow rapidly. Image retrieval technology has become one of the hot spots, and especially Content Based on Image Retrieval(CBIR) technology is widely concerned.CBIR extracts low-level features to represent the image, such as color features, texture features, shape features, and calculate similarities between images by feature vectors to establish the visual links. If a user is looking at an image, other similar images also may cause the user’s interest, which is so-called the mechanism of visual link. The random walk model is an abstract concept model, which is mainly set up for users who have the behavior of browsing the webpage. And many link analysis algorithms are based on the random walk model.If the images and visual links between images are seen as the points on the graph and the transition probabilities of the random walk, the problem of content based image retrieval can be considered from the view of the random walk model.At present, several famous domestic and international search engines input keywords and rank search results by relevance in accordance with the query. The bigger the correlation, the higher the arrangement, which makes the images repeat on the top and gives the user a limited vision. When the user’s query intent is fuzzy or the keyword has multiple semantics, the user may have to click on more webpage to find the image information. Thus we should put emphasis on the above problems and the main research studies are as follows:Firstly, a new algorithm named Two Random Walks(TRW) is proposed to be applied to CBIR, which expands the algorithm called Rank Compete to the structure of manifold. When we only are interested the correlation, compared with Manifold-Ranking Based Image Retrieval(MRBIR), the algorithm can reach higher precision.Secondly, this paper proposes a algorithm named Adaptive Multiple Random Walks(AMRW) to deal with the multiple semantics. The algorithm firstly expands two random walks to multiple random walks, and utilizes the spectral clustering of neighbor propagation to improve the algorithm. The algorithm adaptively selects the representative images to be applied to the diversity image retrieval and presents to the user a diversity vision.