Research On Image Matching Method Based On Image Features

The technology of image matching plays an important role in image processing. It involves many fields of knowledge such as image preprocessing, image acquisition, feature extraction and so on. It closely integrates with many methods of computer vision, numerical calculation and etc... In this thesis, matching image should be pretreatmented and then image features are to be extracted in two ways. Finally, two images thus obtained are to be matched by the fast approximate nearest neighbor search. Compared with classical algorithm, this algorithm proposed in this thesis has some advantages over some other such as in matching accuracy and some main works done in this thesis are briefly described as follows:1. A new non-local median algorithms is proposed to solve the problem of image denoising. By analyzing the advantages and disadvantages of existing image denoising algorithm, we find that the denoising performance of Non-Local Means (NLM) can be improved at large noise levels by replacing the mean by the Euclidean median. Experimental results show that the proposed algorithm has more advantages in de-noising performance and other key performance.2. We consider two aspects to solve the problem of feature point detection. One is based on profile curve and the other is based on gray information.(1) A dominant points detection algorithm is proposed based on adaptive ROS (Region Of Support) to solve the problem of feature points detection based on profile curve. Considering the single measurement technology of ROS at present, the first step is to propose an adaptive monotone detection method. Then, three scanning algorithm is also proposed for the error problem of the curvature caused by the non-maximum suppression, the method achieves a high degree of robust and improves the detection accuracy. And, the algorithm needs no priori parameters, has high running efficiency. And the complexity of the algorithm is only O(N2/2).(2) This thesis introduces the scale invariant feature transform algorithm firstly to solve the problem of feature points detection based on gray-scale image. Then we propose a new algorithm which detect and describe2D features in a nonlinear scale space. The nonlinear scale space is builted using efficient Additive Operator Splitting (AOS) techniques and variable conductance diffusion.Finally, we can get matching results of proposed algorithm in the experimental analysis section. We analysis the feasibility of two algorithms in several ways with the fast approximate nearest neighbor search in this thesis. Experiment result show that detection algorithm based on the nonlinear scale space features has higher detection accuracy in the light changes, scale changes, etc..