高分三号顺轨影像快速匹配并行计算方法

A Parallel Algorithm for Fast Matching of Along-Track Images of GF-3

针对星载SAR影像匹配中影像文件大、重叠区域小等导致匹配效率低问题,该文提出一种基于重叠区域分块匹配策略的SAR-SIFT特征匹配方法,并采用Fork/Join同步并行模式处理金字塔影像以提高特征匹配效率,在特征点提取和特征描述子构建过程中采用OpenMP在CPU端进行并行计算,在GPU端进行逐像元响应函数的并行计算。为验证该方法的有效性,采用2021年覆盖云南省漾濞县“5·21”地震震区的5景高分三号顺轨影像(7104×5650)开展实验,结果表明:在5景高分三号影像匹配中,对实验影像按6×6分块,采用CPU八线程+GPU并行计算方案时效果最佳,加速比为45.15,处理时间从3683.9 s减少至81.6 s,与分块匹配策略、OpenMP多线程和减层降维3种加速方法相比,该方法效率最高,加速比为8.90,可为提高大区域顺轨影像匹配性能提供新的策略和思路。

Aiming at the problems of low matching efficiency caused by large image and small overlapping area in spaceborne SAR image matching,this paper proposes a SAR-SIFT feature matching method based on overlapping area block matching strategy,and uses Fork/Join synchronous parallel mode to process pyramid images to improve feature matching efficiency.OpenMP is used for parallel calculation on CPU in the process of feature point extraction and feature descriptor construction,and the parallel calculation of pixel by pixel response function is carried out on GPU.In order to verify the effectiveness of this method,5 along-track images of GF-3(7104×5650)covering the"5.21"earthquake area in Yangbi County,Yunnan Province in 2021 were used for experiments.The results show that the effect is the best when the image is pressed as 6×6 blocks and using CPU eight threads+GPU parallel computing,with the acceleration ratio of 45.15,and the processing time is reduced from 3683.9 s to 81.6 s.Compared with the three acceleration methods of block matching strategy,OpenMP multithreading and layer reduction and dimensionality reduction,the matching time of this method is the shortest and the acceleration ratio is 8.90.The method proposed in this paper provides a new strategy and idea for improving the matching performance of along-track images in large area.