Improved Differential Evolution Algorithm And SAR Image Registration Based On Edge Points

Differential evolution algorithm is an intelligent optimization algorithm. Due to its simple principle and less controlled parameters, it has been widely used in several areas such as machine intelligence and pattern recognition, and has achieved better results. But there still exist disadvantages that the contradiction between the rate and the precision of convergence and the problem of premature. Therefore, the research of how to improve the accuracy and the convergence speed is very important.Synthetic Aperture Radar can acquire high-resolution images and the obtained data are free from environment and other factors, so it has been widely used in many fields. Since the SAR images at different times on the same area of ground are exist some geometric deviation, it is required for image registration techniques to adjust the images in the same region at different phases for performing following operation, such as image change detection, image fusion and so on. Therefore, SAR image registration is the prerequisite and basis for these subsequent processing, and the corresponding research has very important significance.This thesis is concerned on the differential evolution algorithm and SAR images registration. The contributions of this thesis are two folds:1. We proposed a differential evolution algorithm combined comprehensive learning particle swarm optimization algorithm and clonal selection algorithm. The method introduces a comprehensive learning strategy to guide the individual of differential evolution algorithm to learn from each other, and choose the best individual for cloning to reduce defects caused by greedy algorithm. The mutation mechanism to perform cloning variation is used to the individual which the fitness value has not been updated for a long time. This operation can not only ensure the algorithm convergence in less generation numbers, but also high accuracy of optimization. The final experimental results demonstrated the effectiveness of this method.2. First, the edge responses value of the Hessian matrix are analyzed to give a better threshold, edge points are extracted by a non-maxima suppression method from the speckle images. Compared with the experimental results of Harris corner detection method, this method is effective on speckle images is proved. Secondly, the initial matching parameters is calculated as coarse registration parameters through searching the best match key points by use of improved particle swarm optimization on discrete space. Finally, fine registration is performed based on gray information image registration.