Research On Improvement Of Multi-spectral Image Registration

Multi-spectral image registration is used to match two or more images that get from the same scene, derived from different time, different sensors or different views of angle. Image registration is the key step in image processing and analyzing. It is requisite for images comparison, data fusion and objects recognition.This thesis mainly researches on improving the accuracy and the speed of multi-spectral image registration especially the visible light and IR registration. Aiming at the problem of the standard mutual information which only uses the pixels intensity but ignores their spatial relationship, leading to poor registration, this paper gives a new algorithm named mutual information combining with neighborhood information. The probability distribution of the pixels is determined by the pixels combining with neighborhood information, and the neighborhood information is endowed with different correlation value according to relation degree to the current pixel.Aiming at the problem of the interpolation algorithm used in the registration which make mutual information get into the local extremity point easily, use an improved partial volume interpolation, expanding the interpolation neighborhood. The algorithm reduces the mutations of the joint entropy effectively and overcomes local maximum in mutual information-based image registration.Aiming at the disadvantage of the simulated annealing, this paper gives an improved simulated annealing algorithm. Firstly use the registration result of the two-point method as simulated annealing algorithm's initial value. Secondly limit the new parameter value in the process of searching. At last add a memory device in the process of searching to get a more optimal value.The experiments prove that the algorithms and methods mentioned in the paper can effectively improve the accuracy and the speed of visible light and IR registration.