面向ONSET实时数据处理的图像选帧GPU技术实现  

Implementation on Image Frame Selection of GPU for ONSET Real-time Data Processing

作　　者：李力 邓辉 李瑧 梅盈 戴伟 杨秋萍[1,3] 王锋 Li Li;Deng Hui;Li Zhen;Mei Ying;Dai Wei;Yang Qiuping;Wang Feng(Key Laboratory of Applications of Computer Technology of the Yunnan Province,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Observatories,Chinese Academy of Sciences,Kunming 650011,China;School of Astronomy&Space Science,Nanjing University,Nanjing 210000,China)

机构地区：[1]昆明理工大学云南省计算机技术应用重点实验室,云南昆明650500 [2]南京大学天文与空间科学学院,江苏南京210000 [3]中国科学院云南天文台,云南昆明650011

出　　处：《天文研究与技术》2018年第2期195-201,共7页Astronomical Research & Technology

基　　金：国家重点基础研究发展计划(973计划)(2016YFE0100300);国家自然科学基金(U1231205,U1531132,U1631129,11403009,11463003);云南省自然科学基金(2017FB001)资助

摘　　要：光学和近红外太阳爆发监测望远镜每天可以获得大量的太阳图像数据,对这些观测数据进行实时选帧处理,一方面可以减轻存储压力,另一方面也可以提高后续图像重建的质量。针对观测过程中的选帧要求,设计并实现了一套基于图形处理器的图像选帧实时处理模块,当前的模块已经实现了平均梯度法和谱比法选帧两种算法的高速并行处理。对模块的实现进行了细致的讨论,并比较了两种选帧方法的加速比。实验表明,该模块运行稳定可靠;从执行效率来看,针对近全日面图像的选帧总体执行时间最快为1.2 s,比原有串行实现提升了7倍;局部面图最快为0.7 s,平均提升了5倍。整体模块的实现与当前性能已经可以满足实时观测与处理的要求。Optical and Near-infrared Solar Eruption Tracer(ONSET)can get a lot of sun image data every day.By real-time frame selection to these observation data,on the one hand can the storage pressure be reduced and the observation time be extended,one the other hand the quality of subsequent image reconstruction can also be improved.In this paper,a set of real-time processing image frame selection module based on GPU is designed and implemented for the frame selection in real-time observation process,and it has achieved high-speed parallel processing of gradient method and spectral ratio method in the module.This paper has discussed the implementation of the module in detail,and compared the speed of the two frame selection methods by experiment.Experimental results have shown that the module is stable and reliable.In the implementation efficiency,the best overall execution time of frame selection for nearly full view image is 1.2 seconds,which is seven times better than the original serial implementation.For partial view image it takes 0.7 seconds,and that makes an average efficiency-increase of five times.Implementation and current performance of the overall module have been able to meet the requirements of real-time observation and processing.

关 键 词：实时选帧 光学和近红外太阳爆发监测望远镜 图形处理器 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]