The Feature Points Detection And Descriptors Calculation For Multispectral Images

In recent years, with the rapid development of computer vision technology and the increasing need for image registration, the technology of image registration is becoming more and more popular, many universities, companies and laboratories are doing a lot of research about it. Multispectral image registration technology is rapidly developed because it’s so important in military target detection, biomedical image analysis or the public security monitoring.The two tasks of image registration are the feature points detection and the descriptors calculation. This thesis has studied the fundamental, environment, advantages and disadvantages of the registration method named GS_SIFT, it also studies the important factors that will affect the results of original method.To make the program work faster, the thesis combines SURF detection methods and GS_SIFT descriptors during the key points detection progress, which make the program much faster. On the other hand, study the factors during the progress of building descriptor, such as expanding the size of the descriptor or changing its histogram of oriented gradient, by which make the original program get better results. What’s more, adjust the main direction of all the key points to solve the error matching problem, which occurs while merging the two temporary descriptors.Finally, this thesis achieves the algorithms above, made statistics of the registration accuracy of a lot of image pairs from data set using different methods, analyzed the results of these methods.The main task of image registration is to align two images taken at different condition, such as taken at different times or from different angles. The results show that, compared to the traditional matching methods, combining SURF with GS_SIFT make the original program faster, changing the size of the descriptor or histogram of oriented gradient can achieve better matching results, and the improvement at the last part also improve the matching accuracy, which has a broad space for development.