Research On Image Matching Algorithm With Large-view

Image matching is a fundamental work in the field of computer vision,and this technique is mainly utilized to determine the transformation correspondence of images which contains the same scene by image analysis and processing technology,offering applications in 3D reconstruction,target recognition,stereo matching,image retrieval,image mosaic and other visual fields.Aiming at the problem of image matching in large-view scene,the extraction and matching of local features,as well as the estimation of the relationship model between images are focused on in this paper.The main work and innovations are as follows:(1)A multi-angle feature point detection algorithm based on the angle transformation theory is proposed to solve the inaccurate matching problem of feature matching upon large-view scene.Firstly,the view space of the input image is constructed by the theory of angle transformation,and the nonlinear scale space of the image in the space is constructed by the nonlinear diffusion filtering algorithm.Then the feature point detection operator based on Hessian matrix is used to extract the feature points in the image scale space.The experimental results show that the proposed feature extraction algorithm has good robustness to the change of view and so on.(2)In order to improve the robustness to different view angles,multi-view descriptors and nearest neighbor matching rule based on multi-view information are proposed.Firstly,the binary description of the feature is extracted in the multi-view space,and then the multi-view descriptors are generated by the changes of the descriptors to implement the initial matching.Finally,a multi-view nearest neighbor matching is proposed based on the similarity of the view angles of the feature to obtain the final matching point set.The experimental results show that the image matching algorithm proposed in this paper has a good performance in robustness for large-view change.(3)In order to obtain the accurate image relation model for practical application,an improved random sampling consistent algorithm(RANSAC)based on the effective regions is proposed.Firstly,with the efficient region theory of(GMS)algorithm,trusted and untrusted regions are generated in a certain proportion to initialize the monotone matrix in a guided way.Then,the obvious error model is quickly discarded by a prior test in the same way.The experimental results show that the proposed algorithm can estimate the monotonicity between images accurately and quickly.