| Image matching is one of the most important research fields in computer vision. Its applications in remote sensing have become increasingly significant as well. By studying the image matching and relevant techniques, this thesis elaborates the principles of various types of image matching operators and analyzes the performances of the operators by applying these operators to many different images.The matching methods can be categorized into intensity-based methods, image feature-based methods and other methods. This thesis studies some commonly used methods in the above two categories. By doing experiments, it proves that the image feature-based matching methods have the characteristics of less time-consuming, high accuracy and good adaptability.Based on the above conclusion, the thesis studies some commonly used image feature matching operators. The performances of the operators are analyzed when the images are under the conditions of image rotation, scaling, affine transformation and noise disturbance by comparing the experiment results. And it comes to the conclusion that the SIFT(Scale Invariant Feature Transform algorithm) operator is robust and unique, and it can adapt to the changes of external environment very well. So, SIFT is used to process the images in this thesis.Based on the in-depth study of the basic principles and the SIFT feature point extraction process, the thesis matches the SIFT feature points in two images by using the K-D tree, a data structure which is usually applied to solve the nearest neighbor problems. By using the random sample consensus algorithm(RANSAC), erroneous matching feature points are deleted and transformation relationships between the images are determined. In order to improve the efficiency of RANSAC algorithm, the disparity gradient is applied to constraint the matched points. According to these transformation relationships, the images are stitched together. Finally, the methods are validated by carrying out some experiments. |
PDF Full Text Request