A Study On Registration Of Multi-modal And Huge-size Images

Image registration is a fundamental task in image processing. It is the basic step for multi-image analysis of the same scene. In this dissertation we make creative research on image registration.We suggest a broad sense concept of image match and image registration. New concept of image registration is more accord with the comprehension of human beings on image match. Then we suggest a new match measure: AM (Alignment Metric), which is derived from the judgment of alignment of a human's vision. Mathematically it is based on variance. The new measure can efficiently evaluate the match degree of two images, with little restriction of the gray level properties. It also supplies strong noise tolerance. New measure can be applied to one-dimensional or multi-dimensional multi-modal image registration. Image registration then changes to the maximization of alignment metric. This is a typical multi-dimensional global parameter optimization problem. We use Powell direction set method and multi-resolution analysis to maximize AM. Many experimental results demonstrate the efficiency and performance of the new method.Aiming at feature point based image registration, we suggest the concepts of broad sense feature point and narrow sense feature point. This extends the feature point from a simplex point to a feature region. Based on these concepts we suggest a new kind of broad sense feature point definement. This kind of broad sense feature points can be detected automatically by use of multi-scale wavelet transform and other feature restrictions. When feature point sets are extracted respectively from the two images, correspondence between the point sets is then established by a two-stage matching algorithm. This matching algorithm is based on the Alignment Metric and RMSE (Root Mean Square Error). Finally an iterative optimization mechanism is used to find the precise registration transform of the two images, and then the deformed image is registered. Many registration experimental results can illustrate the performance of this method. We also study on the feature and deformations of remote sensing images. Aiming at infinite length images with local distortions, we suggest a sequential stream-like registration method. Only a very simple manual assistant is needed to startup the subsequent automatic process. The precondition of this method is the continuity and smoothness of local distortion alteration. By setting up grids on the image, broad sense feature points are automatically detected in the reference image and matched with the deformed image. This process can spread to the whole image to evenly establish control points with enough density. Then the local weighted mean method is used to register the image distortions. This method can not only successively register the local geometric distortions but also fit for the infinite length image registration. Images are registered and exported with sequential subsections. Experiments with simulated data demonstrate the efficiency of this method.