Study Of Remote Sensing Image Parallel Processing Algorithms Based On GPU And Optimization Techniques

With the rapid development of remote sensing technology, remote sensing images have been had three features which are high temporal resolution, high spatial resolution and high spectral resolution. They make remote sensing image data showing mass growth rapidly. At the same time, the demanding of the accuracy of real-world applications is ever increasing, so that the remote sensing image processing algorithm has also been toward to more complex computing and more precise. In practical applications, some applications require real-time or near real-time processing speed. Therefore, the processing speed of remote sensing image data is facing enormous challenges.In the high-performance computing area, CPU-GPU hybrid execution model is latest research result at heterogeneous architecture direction. GPU (Graphic Process Unit) has many advantages such as many calculation units. GPU is better in computing processing power than CPU, but not as good as in the logic. CPU-GPU is a new processing mode which is combined with the CPU and GPU respective advantages.The accelerated research of remote sensing image processing has been focused on the use of parallel processing on multi-core CPU. In this paper, we combined with this new heterogeneous execution mode of CPU-GPU and remote sensing image processing speed issues, selecting the most two important remote sensing image processing steps of registration and fusion to study the parallel processing algorithms and optimization techniques. On the one hand, by studying the common parallel model and reasonable division mode of the data, we make the traditional serial processing program under the premise to efficient implementation on the GPU. On the other hand, combining the CPU-GPU architecture features, we try to study the particular optimization techniques of remote sensing. Specifically, the primary work and contributions of this paper are as follows.1. The CPU-GPU heterogeneous execution mode is studied in detail. The new CPU-GPU heterogeneous mode has a variety of execution platform. However, comparatively speaking, NVIDA's GPU programming language is extension of C language, which makes programmers to understand it quickly and plays the GPU execution performance fully. Therefore, this article selected the NVIDIA's GPU as the execution platform and the NVIDA's CUDA programming language as the development tools.2. The theory of remote sensing image registration and fusion are studied in detail. The global registration algorithm based on wavelet decomposition has high accuracy, but it takes longer time. This article selected two registration algorithms with wavelet decomposition that their similarity measures are based on correlation coefficient and mutual information respectively to design and implementation their parallel processing algorithms. There are many types of fusion algorithm, and each algorithm has advantage in certain area. For example, if one is time-consuming, it may be less accuracy, and if one is high precision, it may be time-consuming. This paper selected the direct fusion, HPF fusion, BROVEY transform fusion, YIQ transform fusion and IHS transform with wavelet enhanced algorithm to design and study their parallel processing. The five algorithms have increasing the accuracy of results and the processing time.3. The GPU high performance parallel processing algorithms of registration and fusion are studied and implemented. This paper combined with the GPU's parallel execution model and features, efficient parallel algorithms are developed through the common parallel model, a reasonable division of the data and the correctness of parallel implementation. The experiment results show that the implementation of parallel algorithms had significantly improved performance, and achieved good parallel speedup.4. The optimization strategies of GPU memory level are studied. GPU memory has multi-level structure. The rational and effective using can further improve execution performance. In this paper, we combined the GPU architecture features and data processing methods to study the storage-class GPU optimization strategies for the application of remote sensing images. At the same time, the scalability and adaptability of these strategies were also studied. The experimental results show that these optimization strategies can be well adapted to the high-performance GPU processing of remote sensing image and provide a reference to other areas.