GPU支持的SAR影像几何校正大规模并行处理

GPU supported massively parallel processing for geometric correction of SAR imagery

目的几何校正(又称地理编码)是合成孔径雷达(SAR)影像处理流程中重要的一个步骤,具有一定的计算复杂度,需要用到几何定位模型。本文针对星载SAR影像,采用有理多项式系数(RPC)定位模型,提出了图形处理器(GPU)支持的几何校正大规模并行处理方法。方法该方法充分利用GPU计算资源强大及几何校正过程中每个像素处理步骤一致的特点,每次导入大量像素至GPU,为每个像素分配一个线程,每个线程执行有理函数计算、投影变换、插值采样等计算复杂度高的步骤,通过优化配置dim Grid和dim Block参数,提升GPU的并行性能。该方法通过分块处理实现SAR影像大幅面处理,且可适用于多个不同分块大小。结果实验结果显示其计算加速比为38 44,为全面客观地分析GPU并行处理的特点,还计算了整体加速比,通过多个实验分析影响整体加速性能的因素,提出大块读写提高I/O性能的优化方法。结论该方法形式简洁,通用性好,可适用于几乎所有的星载SAR影像、不同的影像幅面;且加速性能明显。

Objective The geometric correction of SAR imagery is an important step in SAR image processing. This process has a certain degree of computational complexity and requires a certain geometric positioning model. A GPU-supported massively parallel processing method is presented for the geometric correction of space-borne SAR imagery. Meth od This method exploits the RPC model. The method takes full advantage of two facts. A GPU has large computational resources, and the processing steps are the same for each pixel in geometric correction. In the course of massively paral- lel processing, a large amount of pixels are imported into the GPU at each time, and one thread is allocated for one pixel. Each thread performs the steps, which include the calculation of rational function, transformation for projection, resampling, and so on, with high computational complexity. The optimal configuration of two parameters, i. e. , dimGrid and dimBlock, improves parallel performance. Large SAR image frames with different sizes can be processed by block partition. Result Experimental results show that the proposed method can achieve computational speedups ranging from 38 to 44. Meanwhile, the speedup for the whole procedure is recorded to analyze the features of the GPU-based parallel compu- ting objectively and thoroughly. The factors affecting the speedup for the whole procedure are discussed according to the results of several experiments, and an optimal approach that reads and writes a large block to promote I/O performance is proposed. Conclusion The straightforward method has broad applicability and can be used for most space-borne SAR sensors and for different image frame sizes. The method also achieves obvious acceleration.