Algorithms For Image Matching And Its Application Based On Structural Feature

Image matching is the key subject in image processing and computer vision field. It is also the base of theory and application of computer vision. Image matching is widely applied in object recognition, remote sensing and remote measuring, auto navigation, virtual reality, medical diagnosis, automatization of production, military affair and many other fields. Images are strongly structural. And graph, which is a important and effective way to describe the feature information of structure, can keep the connection between regions. Therefore, it is received increasing attention that the structural feature of images is described by graph and that the image feature matching is researched by using graph matching. It becomes the research focus in the pattern recognition at present. It is addressed in this dissertation using the graph to describe the structural connection between the image feature points and researching on the the image structure matching in different condition. The main research works and achievements are outlined as follow:1,An algorithm for image matching in different illumination conditions is proposed, which combines the color vector with singular value decomposition (SVD). Firstly, the existing algorithms for structure matching using SVD are analysed, in which the matching performance for images with large illumination change may be degraded. From the viewpoint of vector correlation in space, the image color vector without the effect of different illumination can be obtained. Secondly, combined with geometric feature of the points in images, the complete weighted graphs describing the structural feature of images are defined. Finally, by performing SVD on the adjacency matrices of the graphs, the correspondence can be obtained by using the relation of feature vectors. Experimental results demonstrate the comparatively high accuracy of the SVD matching algorithm in different illumination conditions.2,An algorithm for image matching combined with SVD and the relaxation is described. Firstly, for the images with large scale, rotation and affine transformation, the complete weighted graphs describing the global structural feature are obtained from the point sets of images. And the initial matching probabilities are computed by using SVD on the Laplace matrices of the graphs. Secondly, the geodesic-intensity histogram with deformation invariance is introduced. And its local similarities are as the compatibility constraints. Finally, the accurate matching probabilities are gained by the relaxation method. So the correspondence of the image points can be acquired from the probabilities. Experimental results show that the proposed algorithm overcomes the limitation of the SVD matching algorithm by the combination of multi-feature and multi-algorithm, improves the matching precision, and promotes the development of the graph method in image matching field.3,An algorithm for non-rigid image matching based on K-nearest Neighbor Graph (K-NN) is presented. Firstly, for the non-rigid image matching with only the image point coordinate information and without any other information, the initial correspondence of the image point sets is acquired by using the shape context method. Secondly, according to the property that the local neighborhood of a point may not change freely due to physical constraints in the non-rigid deformation, the K-NN graphs denoting the local structure information of point sets are constructed. And then the error matching points and outliers are eliminated by making use of the structural differnence in K-NN graphs. Thirdly, the parameters of the Thin Plate Spline (TPS) deformation model between the images are estimated, by which the image point sets are set closer each other. Finally, the matching point pairs in the non-rigid images can be gained by iteration. Experiments indicate that the algorithm with high robustness improves the matching precision and the accuracy of the TPS model parameters between the images.4,Enlightened by the above algorithm, an algorithm for image matching and mosaicking based on minimum spanning tree and TPS transformation model is proposed. Firstly, The SIFT (scale invariable feature transformation) feature points are detected, which are invariable to the noise and the tranformation of scale, rotation and etc.. Secondly, the structure of the feature points is described by using the minimum spanning tree, and the Laplace matrices are obtained with the distance between the SIFT vectors as the weights. Then the correspondence between feature points can be acquied according to the SVD method. Thirdly, the parameters of the Thin Plate Spline (TPS) deformation model between the images are estimated. And by using the iteration method, the final matching point pairs and accurate TPS transformation parameters are gained. Experimental results demonstrate that the proposed algorithm with the higher accuracy than the SIFT algorithm simultaneously solves the two problems of matching and transformation, and that the images are mosaicked effectively.