Research On GPU Based Ultra Wideband SAR Real-time Imaging Technique

In view of the huge computation of conventional BackProjection(BP), this paper integrates two kinds of technology, multi-core parallel on hardware and aperture hierarchical optimization method, to realize real-time imaging with high resolution for Ultra Wideband Synthetic Aperture Radar(UWB-SAR), which is important in theoretic and engineering. The main work is as follows:At first, echo model of striped SAR with stepped frequency signal is established, and the processing of range compression and azimuth compression for BP algorithm is analyzed; then, to implement parallel algorithm on Graphic Processing Units(GPUs), GPU's kernel model and CUDA programming model are introduced in detail; finally, it is summarized that multi-core parallel, storage resource allocation and the storage accessing influence the efficiency of the algorithm, which is the foundation of subsequent algorithm optimization.An optimized parallel BP algorithm implemented on GPU is proposed. Three kinds of optimization methods for the processing of azimuth compression are described in detail: the grid BP algorithm is introduced to reduce the number of multi-threads accessing the original data, texture memory is used to speed up the efficiency of interpolation calculation, and the sub-image storage format is designed reasonably to avoid the access conflict of shared memory, thus the imaging efficiency is improved significantly by making full use of multi-core parallel computing power of GPU.A parallel Factorized Fast BackProjection(FFBP) algorithm implemented on GPU is proposed. The fulfillment Process of FFBP is designed, the acceleration of FFBP at the algorithm level is analyzed, and parallel method is realized with the GPU. The simulations reveal that the imaging efficiency can be dramatically improved compared with the BP algorithm using GPU.The partitioned methods of the above two algorithms is designed for echo data are huge and storage resources on GPUs are limited. Properly focused images are formed and 70× and 232× speedups are obtained using above two methods over BP algorithm implemented on a single-threaded CPU, which verify the effectiveness and veracity of the proposed approach.