Research On Image Retrieval Based On Manifold Learning

Image retrieval is the basic demand in the age of Internet. Content Based Image Retrieval (CBIR) is a hot research direction in the field of image retrieval. CBIR takes advantage of the low-level visual features of images to contrast and retrieve images. It has the characteristics of easy automation and intelligent. At the same time, it conquers the problems in the field of image retrieval based on keywords. For examples, keywords description is not accurate and the retrieval efficiency is not high. CBIR has been widely used in image retrieval of Internet, face recognition and so on.The main parts of the traditional CBIR are feature extraction, dimensionality reduction, similarity search results comparison, classification. There are two serious problems of "dimensionality curse" and "semantic gap" in CBIR. Dimensionality reduction is the main tool of solving "dimensionality curse". And against with "semantic gap" problem, the mainstream in CBIR is using Relevance Feedback (RF) technique. It utilizes the subjective cognitive ability of humans to help improving the accuracy. This paper researches both problems in CBIR system deeply. The main work in this paper is as follows:(l)We analyze the development status of domestic and international in the field of CBIR and discuss existing problems. Thereto, we introduce the definition, background and reasons of "dimensionality curse" and "semantic gap" in detail.(2) The main technology of the CBIR system is described in detail. Thereto, the low-level visual features of images and classification algorithms used for retrieval results are briefly introduced as background knowledge. Dimensionality reduction algorithms, RF and ranking of retrieval results related to the content of this paper are described in details. Meanwhile, the main problems in those techniques are introduced.(3) Manifold learning is a hot topic in dimensionality reduction. This paper reviews the history of manifold learning and introduces some algorithms based on graph structure like Local Linear Embedding (LLE), Local Tangent Space Alignment (LTSA). Moreover, against with the local high curvature in high dimension space, this paper proposes a Wrap Linear Local Tangent Space Alignment (WLLTSA) with angle measurement. WLLTSA sovles the phenomenon of local high curvature in high dimensional space successfully. A lot of experiments have been done to test the performance of WLLTSA.(4) Ranking of retrieval results is the key part of CBIR system with RF. The performance of ranking affects the accuracy and performance of image retrieval directly. This paper makes a brief introduction to the current common similarity measure method. Based on Manifold Ranking (MR) and LRGA proposed by Nie et al. this paper proposes a ULRGA to improve the performance in ranking of retrieval results. A lot of experiments have been done to test the performance of ULRGA.