The Research And Implementation Of Multispectual Image Registration

Image registration is a fundamental problem in many fields. Finding out the transformation between two or more pictures in space, and transforming them according to the transformation is the basic target of image registration which is the precondition of image stitching, fusion and reestablishment. The theoretical basis of multispectral images is the spectral characteristics of different objects which have different characteristics of light assimilation, reflection and radiation. Different requests in different fields, different image data and different image equipment make the multispectral image registration more complex. In the process of image registration, the operation of feature extraction, descriptor and matching make the registration harder and harder which make the research have important theoretical and practical significance.The image registration algorithm based on mutual information and feature which is named “sift” are both researched in this paper. The algorithm based on features is chosen to be used in the multispectral image registration of 31 channels with the reason that the speed of the method based on mutual information is too much slow. The experiments showed that the method based on features could increase the speed and shorten the time with keeping the accuracy of registration at the same time.Considering the registration speed, precision, much image noise and low signal noise ratio, a simplified method with feature detecting and the descriptor of “sift” is proposed for the registration of multispectral images. To solve the problems of complex and large amount of calculation in the “sift” method, the feature detecting operator which can calculate the feature calculation matrix with filtering five times for an image is proposed. In addition, different thresholds can be chosen for different kinds of images which are named “adaptivity” for the registration method, and then a square filter window is used to select features. Finally, the selected features are used for matching with their sift feature descriptor. By the contrast with the sift method, the experiment results showed that this method could shorten the experiment time and improve the efficiency with keeping the accuracy of registration. The adaptation of this algorithm on multispectral images with noise is also proved in the experiment.The image registration develops with the direction of automation and utility. The image between different spectrums which separate with different bands is the emphasis of image registration. The future research will come in two aspects: firstly, we can try to find out the same features in different images with the same coordinate system; secondly, the technology of video frames can be introduced in the image registration which can avoid the repetitive operation in the registration between two images.