Research On Algorithms For Images Matching And Image Segmentation Based On Graph Spectra Theory

The techniques of images matching and image segmentation are both important research areas in computer vision field, which are widely applied in industrial, agriculture, object recognition, remote sensing, biomedicine, military affair and many other fields. Recently, the algorithms for images matching and image segmentation based on graph spectra theory are the most popular research subjects in the world. In the algorithm, pixels in images are processed directly, and the solution becomes rather simply. The algorithm reduces the complexity effectively. So, the research on the images matching and image segmentation based on graph spectra theory has not only important theory meaning but wide perspective of application.The systematic research on the algorithm of images matching and image segmentation based on graph spectra theory is addressed in this dissertation, which includes images matching based on adjacency spectrum, images matching based on Laplace spectrum, images matching based on Laplace spectrum with probabilistic relaxation, image segmentation based on graph spectrum. The main research works and achievements are outlined as follow:1. A novel method for performing point-feature correspondence based on the color and adjacency spectrum is proposed in this paper. First of all, color imformation (characteristic of HSV space or partial accumulation histogram of the hue) and geometric characteristic of the features of both images are mixed into a correspondence strength matrix. Then, the correspondence strength matrix is decomposed by the singular value decomposion (SVD), a relation matrix that denotes the matching degree among feature points is constructed by the result of the decomposition. Finally, the matching feature points of the two images are obtained according to the relation matrix. Experimental results indicate that the algorithm in the paper have the higher precision both of planar matching and of stereo matching for real images.2. One algorithm based on color gradient and Laplace spectrum for image features matching is proposed in this paper. First of all, color gradient information and geometric characteristic of the features of both images are mixed into Laplace matrices. Then the matrices are decomposed by the singular value decomposion (SVD), a relation matrix that denotes the matching degree among feature points is constructed by the result of the decomposition. Finally, the matching feature points of the two images are obtained according to the relation matrix. Experimental results indicate that the algorithm in the paper have the higher matching precision.3. An algorithm of point correspondence based on Laplace spectra of graphs with probabilistic relaxation is proposed in this paper. Given two feature points sets, we define Laplace matrices respectively, analysis the eigenvalues and eigenvectors of the matrices, and obtain the initial correspondence probabilities. Then; the final matching results are acquired by using the method of probabilistic relaxation. Experimental results show that our method possesses comparatively high accuracy.4. An approach for image segmentation by k-means and Normalized Cut is suggested in this paper. The approach uses Normalized Cut to segment between regions after k-means, and produces the final segmented images, k-means focuses On local variations of an image, while Normalized Cut can extract a global property of an image. The experiment shows good results of segmentation.