GPU辅助的希尔伯特变换轮廓术

GPU assisted Hilbert transform profilometry

在希尔伯特变换轮廓术中,相位计算速度的快慢直接影响光学3维测量速度。由于相位计算是在每幅图像的每个像素上分别进行的,每个待处理的像素和已计算的像素之间不存在依赖性,因此具有极高的计算密度,可以通过多线程运行相同的程序进行并行计算。统一架构的图形处理器(GPU)具有强大的数值并行计算能力,可以在GPU上实现并行计算。在分析相位计算的特点后,在GPU上进行相位的并行计算,实现了CPU与GPU的协同工作,解决了在CPU上相位计算方法速度较慢的问题。实验结果表明,经GPU加速后,在相位计算的质量相同的情况下,计算速度有了较大的提高。

The speed of the phase computation affects the optical three dimensional measurement speed in the Hilbert transform profilometry. Because the pixels which still need to be processed and the already processed pixels are not interdependent, the phase of each pixel can be calculated separately. Therefore, the same program can be executed on multiple threads in parallel with high arithmetic intensity. This allows the phase computation done on the GPU with the help of the powerful unified architecture graphics processor parallel computational capability. This paper analyzes the characteristics of the phase computation, and then implements the code for allowing the CPU with GPU working together to ~olve the problem of the low efficiency of the phase computation on CPU. Experiments show that by the G PU acceleration the computation speed has been greatly improved with the same quality of the phase computation.