Fast Image Matching Algorithms Based On New Feature Description

Image matching, as a key technology of computer vision, has been widely used in the fields of 3D reconstruction, image mosaicking and target recognition. The research focus of the image matching is to solve the robustness and speed problems under the conditions of rotation, different scale, lighting, etc. With the development of computer technology and the progress of image matching algorithms, the speed of image matching has made great improvements, but is not enough in some cases with higher real-time requirements, e.g., the missile terminal guidance. As such, this thesis proposes three kinds of fast image matching algorithms as follows:(1) Proposes a fast image matching algorithm based on feedback and task decomposition. The algorithm is based on selective description of feature points. It firstly divides the target image into several sub-areas, and then takes a certain proportion of feature points within each area to describe and match. The matching process will stop if the current result meets the matching requirements; otherwise, according to the previous matching situation of each area, the algorithm takes additional feature points to describe and match based on certain criteria until no feature points are left. The experimental results show that the algorithm can greatly reduce the amount of calculation and speed up of the image matching when combining with feature points based matching algorithms.(2) Proposes a fast image matching algorithm based on triangle feature description. The algorithm firstly extracts corner features of image, and then uses the Delaunay triangulation algorithm to form the triangulation network from the feature points. Next, each triangle in the triangulation network is described using both the edge-angle features and the regional characteristics, which generates a 36 dimensional feature descriptor. By calculating the ratio of the nearest and second-nearest distance, the matching is successful if the ratio is less than a certain threshold. Finally, the proposed algorithm uses RANSAC (random sample consensus) algorithm to eliminate mismatches. The test results using Mikolajczyk image library show that the proposed algorithm can improve the speed of image matching.(3) Proposes a fast line segment matching algorithm based on single endpoint descriptions. The basic idea of the algorithm is using the line endpoints as the feature points. The algorithm proposes to use the direction of the line segment as the reference direction of the feature points, and uses the gradient information of neighborhood to describe the feature points. Finally, the proposed algorithm matches feature points and eliminates mismatches, and determines the line segments matching relationship via line endpoints matching relationship. The experimental results show that the algorithm can improve the speed of the line segment matching algorithm to a certain extent and solve the validation problem of line segment matching.