Image Matching Algorithms Based On Improved Scale Invariant Feature Transform

Image matching is a very important problem in image processing and computer vision. Experts and scholars of various countries have done research for a lot of image matching algorithm. At present, it has been widely applied in many fields, such as image stitching, medical image analysis, target detection, biometrics and other fields. These algorithms are divided into four categories: gray-based, feature-based, template-based and transform domain-based. Since the feature-based matching algorithms are robust to various changes and more flexible, they are the most widely used. In this thesis, feature-based algorithms are introduced, which the most widely used SIFT algorithm is studied and researched in-depth. The shortcomings of SIFT algorithm are analyzed and the corresponding improvement algorithms are proposed. The main research works of this thesis are as follows:1. Aiming at the problems of poor anti-affine and large computation amount of Scale Invariant Feature Transform(SIFT) algorithm for image ma tching, a RadonSIFT method of Hessian-Affine iterative is proposed. Firstly, this method uses SIFT feature detector to extract initial feature points, and then a Hessian-Affine iterative algorithm is used to estimate the affine regions, so that the extrac ted elliptical area meets the requirements of affine invariant. Secondly, the extracted elliptical area is normalized to a circular area. The main orientation is determined and a series of beelines are made in the circular area. Image Radon transform integ ral values on the beelines are adopted as feature vector descriptors. So the robustness for viewing angle changes and the speed of feature matching are improved.2. Aiming at problems of large computation and poor real-time of Scale Invariant Feature Transform(SIFT) algorithm for image matching, a matching algorithm based on Local Binary Patterns(LBP) and Graph Transformation Matching(GTM) is proposed. Firstly, this algorithm uses feature detector of SIFT to extract initial feature points. Secondly, new LBP algorithm is used to produce feature vectors of 29 dimensions for the feature region to reduce complexity of descriptors. Euclidean distance is the measure criterion and initial matching is done. Finally, GTM is adopted to eliminate mismatching points. So the proposed algorithm not only improves accuracy, but also reduces calculated amount.3. Aiming at the problems of large computation, poor real-time and mismatching affected matching accuracy of Scale Invariant Feature Transform(SIFT) algorithm for image mosaic, a Radon-SIFT method based on adaptive non-maxima suppression(ANMS) is proposed. Firstly, this method uses feature detector of SIFT to extract initial feature points and then the initial feature points are preferred by ANMS, so the feature point set of uniform distribution is obtained. Secondly, a series of lines are made in the feature area. Image Radon transform integral values on the lines are adopted as feature vector descriptors and features are matched by measure criterion as Euclidean distance. Random sample consistency(RANSAC) algorithm is used to eliminate mismatching points. So the computation speed and correct matching rate are improved.