摘要
为实现高光谱影像数据快速降维,基于nVidia的图像处理单元(graphic processing unit,GPU)研究最大噪声分数变换(Maximum Noise Fraction Rotation,MNF Rotation)降维算法的并行设计与优化,通过对加速热点并行优化,择优整合,设计并实现基于CUBLAS(CUDA Basic Linear Algebra Subprograms)库的MNF-L(MNF-on-Library)算法和基于CPU/GPU异构系统的MNF-C(MNF-on-CUDA)算法.实验结果显示MNF-L算法加速11.5~60.6倍不等,MNF-C算法加速效果最好,加速46.5~92.9倍不等.研究结果表明了GPU在高光谱影像线性降维领域的巨大优势.
To rapidly reduce the huge dimensions of hyperspectral image, this paper investigated the design and optimization of the parallel Maximum Noise Fraction （MNF） Rotation algorithm based on nVidia graphic processing units （GPUs）. In particular, a MNF-L （MNF-on-Library） algorithm based on the CUBLAS library functions and a MNF-C（MNF-on-CUDA） algorithm on the CPU/GPU heterogeneous system was presented by designing mapping schemes and parallel optimizing strategies. Experiment result shows that the MNF-L algorithm can obtain the speedups between 11.5 and 60.6, and the MNF-C algorithm can get the speedups between 46.5 and 92.9. Therefore, GPUs have a great advantage in reducing dimensions of hyperspectral images.
作者
周海芳
高畅
方民权
ZHOU Haifang GAO Chang FANG Minquan(School of Computer, National University of Defense Technology, Changsha 410073, China)
出处
《湖南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2017年第4期147-156,共10页
Journal of Hunan University:Natural Sciences
基金
国家自然科学基金资助项目(61272146)
国防科学技术大学优秀研究生创新资助项目(B151101)~~
关键词
图像处理单元
GPU性能优化
高光谱影像降维
最大噪声分数变换
协方差矩阵计算
graphic processing unit
performance optimization for GPU
dimensionality reduction of hyperspectral image
maximum noise fraction rotation
covariance matrix calculation
