Research On Methods For Features Matching Based On Spectral Graph Theory

Feature matching is a basic problem in computer vision and pattern recognition, which is an essential process in the specific application of three dimensional reconstruction, image registration, image retrieval, object recognition and classification. As a kind of representative method of feature matching, the basic idea of method for matching based on spectral graph theory is to transform feature matching problem into graph matching problem and use matrix to describe the relationship of features. The matching results will be got by analyzing the spectral characters of this matrix. The method for matching based on spectral graph theory provides a good solution channel for feature matching problems in the case of complex transformation. Its calculation is simple and can overcome combination explosion problem in the graph matching effectively. In this paper, the related research work has been done around feature matching based on spectral graph theory. The main research work and achievements are outlined as follow:1. The systematic explore on the algorithm of feature matching based on spectral graph theory is addressed in this dissertation. It is based on the summary of model algorithms. According to the mathematical theory, the research on the principle and essence of the spectral methods in the different matching algorithms is addressed.2. This work analyzed two typical methods in the past, proposed a novel approach of spectral graph matching based on new adjacency matrix. This new method constructs an affinity matrix which not only considers the geometrical similarity of the same image and the different images, but also adds the texture similarity weight factor of features. Experimental results show that this method does well in the condition of rotation, translation, scale transformation and distortion of images and better than the above-mentioned methods. 3. A method of edge similarity weighted is addressed. The basic idea is offering a big weight coefficient to the edges of more discriminating and reducing the weight coefficient of bad discrimination edges. Like this, the spectral characters of affinity matrix can reflect the matching relation of features more reliable. Experimental results show that this weighted method can improve the property of spectral matching methods effectively.