摘要
为解决高质量的轮廓提取算法计算复杂、实时性差的问题,基于GPU并行计算架构提出了一种针对高质量的轮廓提取算法——Pb(probability boundary,概率轮廓)提取算法的高效并行计算方法。重点讨论了如何利用多计算单元加速计算最耗时的梯度计算部分。详细介绍了多方向直方图并行统计机制及χ2并行计算中访存冲突避免机制。对比实验表明,在GPU上基于该并行方法的轮廓提取相比传统CPU方式具有明显加速效果,且随着图像分辨率变大,加速效果更加明显,例如图像大小为1024×1024时可获得160倍的加速;此外,基于伯克利标准测试集验证了该并行方法可保持原有算法的计算准确度。为大规模图像数据智能分析中的轮廓提取提供了快速、实时的计算方法。
In order to solve the problem of high computational complexity and poor real-time performance in high-quality boundary detection algorithms, this paper proposed a high-efficient parallel computing method based on GPUs for the Pb algorithm, one high-quality boundary detection method. This paper paid more attention to accelerating gradient computation, which was the bottleneck of boundary detection computation. The method to process histogram statistics in parallel were discussed in more detail, as well as the method to avoid bank conflicts on shared memory when computing Х^2 differences in parallel. Experimental results show that the original CPU-based Pb algorithm is accelerated significantly by deploying the parallelized method on a GPU. The acceleration effect becomes more distinguished with increasing image sizes. Take 1024 ×1024 as an instance, it obtains a 160x improvement by employing GPU-based optimized method. Moreover, it also shows that the parallel computing method is able to obtain the same detection accuracy with that of original when Berkeley datasets are used as the test bench. This paper provides a reference method for high speed and real-time analyzing of high volume image data.
出处
《计算机应用研究》
CSCD
北大核心
2015年第2期630-634,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(61170121)
高等学校学科创新引智计划资助项目(B12018)
