The Implementation And Parallelization Of SAR Image Processing Algorithm Based On GPU

Remote sensing image processing is of important practical significance. Because of the big data to be processed, using CPU to process the remote sensing images can not satisfy the real time request. GPU can largely improve the speed of remote sensing image processing because of its inherent characteristics of multi-kernel. However, the complexity of multithread programming limits the application of GPU in remote sensing image processing largely. So It is very siginifivant to study multithread programming and the parallel planning of remote sensing image processing algorithms based on GPU.In this paper we mainly study image registration, change detection and image fusion algorithms which are very commonly used in remote sensing image processing, we develop a remote sensing image processing algorithm with a CPU plus GPU heterogeneous implementation. For the proposed algorithm, CPU is mainly responsible for algorithm logic control because of its good logic controlling characteristics, while GPU is mainly responsible for the data processing section because of its multi-kernel characteristices. For the implementation of the image registration algorithm of GPU parallelization, by combining experiments and the test data, we analyze the advantage of multi-kernal processor in image registration in terms of speed, price, power consumption and convenience, and we slove the data access conflicts of obtaining the image histogram in image registration for the Pyramid of mutual information. For the parallelization of the two algorithms of change detections based on Markov Random Fields and wavelet transform, we mainly solve the parallelization implementation problem of parellel planning and image convolution, then we design and implement a scheme for the parallization of Markov Random Fields algorithm and wavelet transform algorithm, the contrast experimental results demonstrate that the realtime performance of change detection is highly improved under the premise in detection precision. For the parallelization implementation of the remote sensing image processing algorithm, we combine the parallel architecture of currently implemented wavelet transform algorithms with the construction of the pyramid in image registration, design and implement the parallelization of the image fusion algorithm based in wavelet transform, measure the time of the implemented parallelization scheme and the scheme on CPU platform, and compare the run time. The experimental results demonstrate that the proposed and implemented parallization scheme in this paper is suitable for the fusion of different kinds of images and the speed is increased by more than 30 times compared to the scheme on CPU platform. Finally, we briefly introduce the application platform of the research result of this paper and propose a scheme to further improve the realtime characteristics of the algorithm.