Study And Implementation Of Parallel Algorithms For Image Registration Based On Mutual Information

Image registration is an important problem in image fusion, image analysis, change detection and target recognition. Applications of image registration are in the domain of military, remote sensed data processing, medical, computer vision and so on. The mutual information, which originates from information theory, is the measure of two random variables statistical relevance, and can be applied in image registration as a similarity measure. Since the mutual information has advantages including no pre-processing, robustness and high automatic processing, it becomes a hot topic in image registration. With the rapid development of sensor and remote sensing platform technology, there are more and more approaches and sorts of the information we can get . As a result , the data sets gotten from the sensor are increasing tremendously. Meanwhile, rapid processing technologies and real time process in many domains become more and more urgent, and the traditional single-processor can't achieve the requirement. However, the key to solve the problem is using parallel processing. Since both the need for image registration and the amount of data to register are growing tremendously, the implementation of automatic image registration methods on high-performance computers needs to be investigated.Nowadays, the current research is focusing on improving the registration quality mostly, but the research of parallel processing on the large number of data sets and the real time request in registration algorithms are in the beginning. So we do some research on parallel algorithms for image registration that could make great progress in the efficiency and the speed of image registration, which can be applied broadly.This paper aims at the research and implementation of parallel algorithms for image registration processing based on mutual information. The contributions and relevant work in the paper are as follows:Firstly, we discuss the mutual information computation methods which are based on Histogram and Parzen window. And then we analyze them and compare them with each other. It shows that the method based on Histogram is coarser, but it is faster than the method based on Parzen window. The Parzen Window method is determined by the window function's form and width. With this method, when samples are enough it is precise. But its cost is great because it has to be added up many times.Secondly, we discuss three typical kinds of similarity measure, which are Mutual Information, Correlation Coefficient and Sum of Square Differences. Then, the performance of these measures is compared and analyzed by computing time, sharpness, sensitivity of noise and effect on multimodal image registration. The results of experiments show that these similarity measure methods have different effect and performance in the different application circumstances, but the mutual information is more precise.Thirdly, a rigid parallel image registration algorithm based on mutual information is presented and implemented. The algorithm uses rigid transform, and it solves the problem of load balancing with image data vertical division, data block parallel inputting. Binary Tree Reduction is used in parallel computing mutual information, which speeds up the registration rate and increases the computation efficiency. The experiment results show that the algorithm has a good speedup, scalability, and high parallel efficiency.Finally, we present a parallel non-rigid image registration algorithm model (based on Basic Function), and then we use the model to parallel a non-rigid algorithm. The parallel technologies applied to the model are parallel inputting the Cross-division of the data, parallel automatic selecting marker, Binary Tree Reduction calculation of the mutual information and parallel outputting the block images. Theoretical analysis shows that the algorithm model has a high parallel efficiency and a relatively strong versatility.