Techniques Of Multispectral Image Registration Based On Mutual Information And Corner

The performance of target recognition is usually limited when using single sensor. In order to get robust results, an effect way is to take advantage of multi-source information to do the work, which needs fusion of multi-spectral images. If the space in satellite or missile is limited which leads to unregistered images of different sensors or images of different time intervals, then we must register them before use. In this thesis, the main contributions are given as follows: Discussed the feasibility and advantage of using mutual information to do multispectral image registration. It uses the structures of images to avoid any hypothesis of the relationship of the gray values. Neither are segmentation and preprocessing needed(such as feature extraction, scene classification). Compared the three interpolation algorithms of nearest neighbor, linear and PV(Partial Volume), experiment results indicated that PV interpolation gave the best performance in terms of the smoothness of mutual information function and the precision of registration. Researched the reason why PV interpolation may lead to wrong registration, and suggested to reduce gray levels to avoid the problem. Researched the optimization theory in image registration based on maximizing mutual information, proposed to use simulate annealing-simplex method to do the global optimization and to use a new cycling method the deal with edge restriction which was compared with impingement method. Experiments showed that registration based on maximizing mutual information using improved simulate annealing-simplex method realized auto-registration of multi-spectral images and reached sub-pixel precision. Due to the inferiority of image registration using maximum mutual information, we researched the registration based on corner points. We used Harris operator in consideration of repetition and information measure. Mutual information was proposed to be the similarity measure of corner points and was compared with unified correlation coefficients and Cluster Reward algorithm. Experiments showed best result when using mutual information. When initial control point pair is obtained, variance check and error check were used to exclude fake control point pairs to get the final ones. Experiment showed high precision and reliability of this method.