摘要
针对传统的图像模式识别算法提出一种GPU加速新方法,结合对CUDA架构的分析,通过充分利用GPU优秀的并行计算能力和高存储器带宽提高图像处理速度.分别对不同大小及不同批数量(单次处理图像数)的图像进行识别处理,并对其进行了多种优化,实验证明相同算法在GPU上的实现与CPU相比处理速度最高提升了600倍左右,达到了平均每幅图优于2 ms的处理速度.此项技术已成功应用于高速铁路扣件在线探测,为高速铁路安全信息在线监测提供了新的有效的手段和方法.
A new acceleration method based on GPU for the traditional pattern recognition algorithm is proposed, and the CUDA architecture is analyzed. The excellent parallel computing capability and high memory bandwidth of GPU are utilized to enhance the image processing speed. Different resolutions and numbers of images are also used to test the algorithm efficiency with a variety of optimizations. Experiments show that this algorithm is so efficient that the CPU-GPU speed-up can reach about 600 times, and the processing time of each frame is reduced to 2 milliseconds. The technology is successfully used to the real-time detection system of high-speed rail fasteners, which provides an efficient measure for the guarantee of high-speed railway safety.
出处
《大连交通大学学报》
CAS
2011年第6期63-67,共5页
Journal of Dalian Jiaotong University
基金
国家863重点计划资助项目(2009AA110300)
关键词
图像模式识别算法
GPU
加速比
orientation field algorithm
graphics processor unit(GPU)
speed-up