基于CUDA架构下的直方图均衡并行算法  

Histogram equalization parallel algorithm based on CUDA architecture

作　　者：肖汉 肖诗洋 孙陆鹏 郭宝云[3] XIAO Han;XIAO Shi-yang;SUN Lu-peng;GUO Bao-yun(School of Information Science and Technology,Zhengzhou Normal University,Zhengzhou 450044,China;School of Civil Engineering,Northeast Forestry University,Harbin 150040,China;School of Civil and Architectural Engineering,Shandong University of Technology,Zibo 255000,China)

机构地区：[1]郑州师范学院信息科学与技术学院,郑州450044 [2]东北林业大学土木工程学院,哈尔滨150040 [3]山东理工大学建筑工程学院,山东淄博255000

出　　处：《桂林理工大学学报》2021年第3期654-663,共10页Journal of Guilin University of Technology

基　　金：国家自然科学基金项目(41701525,61572444);山东省重点研发计划项目(2018GGX106002);山东省自然科学基金项目(ZR2017LD002)。

摘　　要：为了提高图像对比度,解决传统的直方图均衡算法处理速度慢的问题,提出了基于统一计算设备架构(CUDA)的直方图均衡图像实时处理加速方案。利用图形处理器(GPU)强大的计算能力和CUDA优化的存储器结构,以加速直方图均衡中的图像灰度级投票、分布概率累加并映射以及图像新的灰度值填充等功能进行运算。在CUDA条件下,对直方图均衡算法进行串-并行分析,分别从粗粒度与细粒度角度进行并行设计,通过实验测定进行了设计参数寻优,获得了线程块设计参数的最优尺寸。结果表明,基于CUDA的图像直方图均衡并行算法的性能相比基于CPU的串行算法和基于开放多处理(OpenMP)并行算法分别获得了61.58和32.00倍的加速比,能够为大规模实时性图像处理系统设计提供参考。In order to improve the image contrast and speed up the slow processing speed of traditional histogram equalization algorithm,a histogram equalization image real-time processing acceleration scheme based on Compute Unified Device Architecture(CUDA)is proposed.The powerful computing power of Graphic Processing Unit(GPU)and the optimized memory structure of CUDA are used to accelerate the operation of image grayscale voting,distribution probability accumulation and mapping,and image new gray value filling in histogram equalization.Under the condition of CUDA,the series-parallel analysis of histogram equalization algorithm is carried out.Parallel design is carried out from the point of coarse-grained and fine-grained respectively.Through the experimental determination,the design parameters are optimized,and the optimal size of the thread block design parameters is obtained.The experimental results show that the performance of the image histogram equalization parallel algorithm based on CUDA is 61.58 times and 32.00 times higher than that of the serial algorithm based on CPU and the parallel algorithm based on Open Multi-Processing(OpenMP),respectively.At the same time,compared with the relevant literatures,it has greater performance advantages,and provide reference for the design of largescale real-time image processing system.

关 键 词：直方图均衡 图像增强 图形处理器 统一计算设备架构 并行算法 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]