基于GPU的小尺寸FFT在实时图像复原中的优化

Optimization on FFT of small size in real-time image restoration based on GPU

为满足跟踪识别系统对图像复原的实时性需求,在图形处理器(GPU)上进行高效实现小尺寸二维FFT的优化策略研究。首先对二维FFT算法进行分析,根据图形处理器的特点,提出基于图形处理器的并行执行模型。基于该模型,从算法的复杂度、跳转指令的数量、共享存储器的访问冲突以及共享存储器的访问延迟及图形处理器的利用效率这4个方面进行优化策略的研究,提出相应的优化方法。在图像复原的实验中,先对基于GPU的小尺寸FFT优化方法与基于CPU的MATLAB传统算法进行计算精度对比,然后基于4种不同尺寸的图像在相同的GPU平台上再与NVIDIA公司提供CUFFT函数库复原算法进行计算效率对比。研究结果表明:该优化方法提供的图像复原算法复原效果好,与MATLAB效果图比较人眼观察不出差异;在计算速率上,提出的优化方法能够在19.6 ms内复原1帧128×128灰度模糊图像,计算速度与直接采用CUFFT函数库算法相比提高约1.8倍。

To meet the real-time demand of image restoration for recognition and tracking system,an optimization research on two-dimensional FFT of small size realized in graphics processor unit(GPU)efficiently was done.An analysis of two-dimensional FFT algorithm was analyzed first.And according to the characteristics of GPU,a parallel execution model based on graphics processor was proposed.Based on this model,the optimization research was done considering the aspects of algorithm complexity,the number of jump instructions,access conflict and access latency of the shared memory,and the utilization efficiency of GPU.And two-dimensional FFT computation of small size was realized in the GPU.In image restoration experiment,comparison on the calculation accuracy of two-dimensional FFT of small size optimization algorithm based on GPU and the traditional algorithm in MATLAB based on CPU was done.And a comparison on the computational efficiency of optimization algorithm proposed and the library function image restoration algorithm of CUFFT provided by NVIDIA Corp in four different sizes based on the same GPU platform was made.The results indicate that this optimization algorithm has excellent recovery performance,and human vision system could not distinguish the difference between the results and the MATLAB demonstrations.And the optimization algorithm can recover a frame of128×128gray fuzzy image within19.6ms,while the computing speed increases1.8 times approximately compared with that using library function of CUFFT directly.