Study Of Parallel Algorithms For Remote Sensing Image Registration Based On GPU And Implement Of Application System

Image registration, stands as an indispensable step in remote sensing algorithm, has been applied in many practical remote sensing applications. As the temporal resolution, spatial resolution and the spectral resolution being improved constantly, the problem of huge expansion size in data of remote sensing image has arisen. At the same time, image registration, which is a typical computation-intensive and memory access intensive procedure, has a relative high computational complexity, thus the traditional serial processing model cannot keep up with the speed requirement of the high-end applications in military and agroforestry. Meanwhile, along with the unceasing improving performance of GPU, GPU general purpose computation has become a hotspot of computational technique providing a promising insight into speeding up the remote sensing image processing.In this paper, contra posing to the two typical methods of region-based and feature-based in image registration, a parallel image registration algorithm and its optimizing strategies are given. Also, we design and implement a parallel processing prototype system applying our algorithm. The main contributions of our work are detailed below:1. The heterogeneous execution model of CPU-GPU is studied. Especially the GPU from nVIDA with CUDA programming model is chosen to implement the CPU-GPU heterogeneous parallel processing algorithm.2. Parallel algorithm for region-based registration based on GPU is researched. After designing a global registration algorithm which uses the iterative refinement method, a GPU parallel computing model suited for the algorithm above is proposed. Also, the optimizing strategies on data loading, thread memory accessing and circulation disintegration are studied and the optimization is verified by the experimental result.3. Parallel algorithm for control-point based matching based on GPU is researched. Searching control points and match parameters is the key step of this kind of image registration method, which involves irregular data access, multiple branches, loop iteration, and the parallel design or optimization is more difficult. Select a control point matching algorithm based on mutual information as the research object, and design two kinds of GPU parallel scheme with the focus on mutual information calculation and least-square matching process on the basis of the data flow analysis. In hard to eliminate iteration related cases, results show that local storage optimization and atomic operation can still get more than 10 times acceleration.5. A remote sensing image parallel processing prototype system based on Web is designed. The system adopts B/S mode with Java, and provides image processing services, and integrates with 12 classes, including the above algorithm research, a total of 49 kinds of remote sensing image parallel processing algorithms. The system also provides a friendly interface and good extensibility.