马秀丹¹, 崔宾阁¹, 钟 勇¹, 张永辉¹, 费 东¹
山东科技大学 计算机科学与工程学院

摘要:
高分辨率遥感图像处理经常面临程序执行时间过长和内存空间不足的问题, 虽然并行计算技术可以提高遥感图像的处理速度, 但是无法降低算法占用的巨大内存空间。为了解决这一问题, 本文提出了一种利用CUDA和内存映射文件的高分辨率遥感图像快速处理方法, 并以K-Means算法为例进行了实现。其中, CUDA技术可以有效利用GPU 强大的并行计算能力, 而内存映射文件技术降低了磁盘I/O速度较慢对算法性能的影响。实验结果表明, 本文方法比传统K-Means聚类算法计算速度提高了30倍左右, 内存使用量降低了90%以上。
关键词:内存映射文件CUDA优化遥感图像聚类
DOI：10.11759/hykx20171011025
分类号:
基金项目:国家自然科学基金项目(41406200); 山东省自然科学基金(ZR2014DQ030)

A fast processing method for high-resolution remote sensing image based on GPU and memory mapping file
MA Xiu-dan,CUI Bin-ge,ZHONG Yong,ZHANG Yong-hui,FEI Dong
Abstract:
High-resolution remote sensing image processing programs often encounter the problems of lengthy execution time and insufficient memory space because of the limitations of computer processor frequency and memory capacity. Parallel computing can enhance remote sensing image processing speed, but is unable to reduce the requirements for memory space. In this paper, we present an approach to simultaneously optimize the time and space efficiencies for image processing programs. This has been previously implemented for the K-Means algorithm. CUDA can effectively take advantage of GPU powerful parallel computing capabilities, and memory mapping file can offset the impact of a slow disk I/O performance on algorithms. Experimental results show that our method is about 30 times faster than the traditional K-Means clustering algorithm, and the memory usage decreased by more than 90%.
Key words:Memory mapping fileCUDAOptimizationRemote sensing imageClustering





PDF全文下载地址:

点我下载PDF