高光谱RX异常检测的多DS P并行化处理技术  

Multi-DSP Parallel Processing Technique of Hyperspectral RX Anomaly Detection

作　　者：郭文记[1] 曾晓茹[1] 赵宝玮[1,2] 明星[1] 张桂峰[1] 吕群波[1] 

机构地区：[1]中国科学院光电研究院,计算光学成像技术重点实验室,北京100094 [2]西安电子科技大学,陕西西安710071

出　　处：《光谱学与光谱分析》2014年第5期1383-1387,共5页Spectroscopy and Spectral Analysis

基　　金：国防基础科研计划项目(B1320110041)资助

摘　　要：为达到高光谱图像数据RX异常检测处理的高速、实时、海量存储等要求,本文提出了一种基于CPCI Express标准总线架构的多DSP高光谱图像并行处理系统的解决方案。系统采用4片DSP共享数据总线和存储器的紧耦合与Link口互联相结合的硬件拓扑架构。在该硬件平台上,针对光谱RX异常检测算法以及光谱图像三维数据的特点,分配各DSP并行处理任务,提出了一种利用图像空间分块计算并求整个图像的均值矩阵与协方差矩阵的4DSP并行处理技术。结果表明,在保证同等探测效果的条件下,采用本文的RX异常检测算法4DSP并行处理技术,可以达到单DSP处理4倍的时间效率,解决了DSP内存容量对大数据量图像处理的限制,并较好的完成了光谱数据的实时处理要求。To satisfy the requirement of high speed ,real-time and mass data storage etc .for RX anomaly detection of hyperspec-tral image data ,the present paper proposes a solution of multi-DSP parallel processing system for hyperspectral image based on CPCI Express standard bus architecture .Hardware topological architecture of the system combines the tight coupling of four DSPs sharing data bus and memory unit with the interconnection of Link ports .On this hardware platform ,by assigning parallel processing task for each DSP in consideration of the spectrum RX anomaly detection algorithm and the feature of 3D data in the spectral image ,a 4DSP parallel processing technique which computes and solves the mean matrix and covariance matrix of the whole image by spatially partitioning the image is proposed .The experiment result shows that ,in the case of equivalent detective effect ,it can reach the time efficiency 4 times higher than single DSP process with the 4-DSP parallel processing technique of RX anomaly detection algorithm proposed by this paper ,which makes a breakthrough in the constraints to the huge data image pro-cessing of DSP’s internal storage capacity ,meanwhile well meeting the demands of the spectral data in real-time processing .

关 键 词：高光谱图像 并行处理 RX异常检测 多DSP FPGA 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]