Study On Fast Image Matching Algorithm Based On Scale Space And Feature Description

The function of image matching is to determine the corresponding relation between two images gained from different times, different angles, different sensors or different shooting conditions, it is indispensable in the target location, visual navigation, target recognition,3D reconstruction, image stitching, etc. In the task of image matching, the matching speed is relatively slow though the matching quality is acceptable. Therefore, this thesis focuses on increasing the speed of image matching algorithm. Since scale space needs to be built in order to ensure the scale invariance of image matching, and the construction of descriptors is crucial for capturing the representative information of an image, this thesis studiesthe scale space and the construction of descriptors in an effort to improve the speed of image matching but ensure the quality of image matching. The main work is as follows:(1) This thesis proposes a fast matching algorithm for large-size images based on image preprocessing. This algorithm firstly analyses the reasons why the matching speed of large size images is slow, and then provides corresponding solutions. Before image matching, we shrink the large size image by a determined ratio using three proposed methods, including the testing method, model-based method and feedback-based method. The experimental results show that the proposed algorithm can not only effectively improve the matching speed of large-size images, but also the matching accuracy. The algorithm is easy to be combined with the existing matching algorithms containing scale changes to further enhance the performance of large-scale images matching.(2) This thesis proposes a fast corner matching algorithm based on binary description FREAK. First, by building a simple Gaussian pyramid-shaped, the scale invariance is incorporated into the corner of FAST. Then, by reducing the overlapping region, this thesis proposes a new way to build FREAK descriptors. A large number of experimental results show that the algorithm can improve the speed of image matching and yield matching of the actual aerial image that SIFT, SURF, ORB and BRISK cannot match.(3) This thesis proposes a fast matching algorithm based on the scale-space processing using polynomial representations and binary describe LDB. Firstly, a continuous sLoG scale space is constructed by using polynomial decomposition, and then blobs are detected in that scale space. Regarding the descriptor building, the differences of descriptors and building speed are expected to be enhanced by adding main directions of descriptors, changing the grid partitioning of feature point neighborhood, and resampling the grids, the construction method of binary string descriptor LDB is improved in. Extensive experimental results show that the matching speed of the proposed algorithm could satisfy real-time requirements, and achieve a good matching performance as well.