Study Of Image Feature Extraction And Matching Algorithms Based On Corner Points

Image feature extraction and matching have been widely used in computer vision. The stability of the image feature extraction is the precondition of image matching. Depending on the type of local feature points extracted in the image, the points can be roughly categorized as blob points and corner points. Due to the influences of the scale, rotation, light, noise and other factors, the actual captured images, e.g, the aerial buildings, are often loss of edge information, so that only a small number of blob features can be extracted on the edges of the image structure. In contrast, corner points are able to maintain the stability of the object edges features. Thus, when this kind of image matching cannot be achieved with the blob features, corner points become an important guarantee to realize the matching.However, the existing algorithms for extracting corner points have shortcomings in terms of the precision of matching. As such, in this thesis, we carry out the study of image feature extraction and matching algorithms based on corner points. The main work is as follows:(1) The thesis studies not only the typical blob algorithms SIFT and SURF, but also the corner algorithms Harris and FAST. By the analysis of the experiment using the building images, the results verify that the performance of corner points detected on the edge structure is better than that of blob points and with a faster detection.(2) The thesis proposes a new image feature extraction and matching algorithm based on the FAST corner points. The scale invariance is introduced into the FAST algorithm to extract the scale invariant corner features. The SIFT algorithm is then combined to build the128dimension feature descriptor with nice performance. For the feature matching, the thesis uses the two-way matching method and the new eliminated false matching algorithm to improve the matching accuracy. Experimental results based on different images datasets show that the proposed algorithm has higher matching precision and improved robustness than SIFT, and can realize the actual aerial image matching that SIFT algorithm failed.(3) The thesis proposes a novel image feature extraction and matching algorithm based on the BRISK corner points. Combining scale spaces of the SIFT algorithm with BRISK algorithm, the thesis puts forward a new method for building the scale spaces. As a result, the BRISK algorithm is exploited to extract the scale invariant corner points. Then, the thesis utilizes the improved feature matching method and the improved elimination algorithm for false matching to realize the correct image matching. The experimental results show that the proposed algorithm is an efficient image matching algorithm, which has higher matching precision than the new algorithm based on the FAST corner points, and also realizes the actual aerial image matching that BRISK and SIFT algorithm cannot achieve the matching.(4) The thesis studies nonlinear image feature extraction and matching algorithms, and implements the nonlinear KAZE algorithm and its improved algorithm. Extensive evaluations are performed with different images datasets to compare the nonlinear algorithms with the state-of-the-art linear algorithms such as SIFT, SURF, ORB, BRISK and FREAK. By carrying out the experiments under different situations,such as scale changes, rotation changes, view changes, fuzzy changes, illumination changes, image compression changes and deformity changes, the results are obtained, indicating that the performance of the nonlinear algorithms are better than that of the linear algorithms. Compared with the KAZE algorithm, the improved KAZE algorithm not only shortens the processing time, but also improves the accuracy of matching.