High Speed SAR Imaging Methods Based On GPU

Synthetic aperture radar(SAR)is an important remote sensing technology,which has all-day,all-weather capabilities and abundant echo information.SAR has great potential for application in many fields,such as military,agriculture and forestry.For current SAR system,the target resolution has reached centimeters,the amount of received echo data is huge and the SAR image is obtained by complex imaging algorithm.Many applications of SAR system have real-time imaging requirements,which lead the high-speed imaging processing becomes a key technology.Since Graphics Processing Unit(GPU)entered the era of unified rendering architecture in2006,its computing performance and storage bandwidth have been greatly improved.This thesis dedicates to accelerate the computing speed of SAR imaging process by using the large-scale parallel architecture of GPU,proposes a high-speed computation method for SAR real-time imaging system.The main research work and contributions of this thesis are shown as follows:1.The changes of hardware architecture are studied from the birth of GPU,detailed analyze the computing performance and the memory hierarchy of Maxwell architecture.According to the problems encountered in the research process,summarize the programming essentials about designing GPU parallel computing functions on the Compute Unified Device Architecture(CUDA)platform.2.The imaging principle of SAR and the general imaging algorithms are introduced.From the engineering point of view,an improved range Doppler(RD)imaging algorithm is presented,which is suitable for parallel computation on GPU.The real echo data are used for imaging test,and a clear imaging result is obtained,that verified the feasibility of the proposed method.3.For further acceleration of SAR imaging process,a parallel computing method of FFT based on Stockham autosort framework and an in-place matrix transposing method are presented.By integrating computing process,the speed of pulse compression has increasedabout twice compared with the current acceleration methods that use cuFFT library.8192×4096 points echo data can be imaged in 68 milliseconds by using proposed methods on GTX 970 GPU.At the same time,the cost of memory space has decreased by 2/3.