基于CUDA的快速人脸检测算法  

Fast Face Detection Algorithm Based on CUDA

作　　者：孙立超[1] 张盛兵[1] 程训焘[1] 张萌[1] 

机构地区：[1]西北工业大学计算机学院,陕西西安710129

出　　处：《计算机与现代化》2013年第8期11-14,共4页Computer and Modernization

基　　金：高等学校博士学科点专项科研基金资助项目(20116102120049);西北工业大学基础研究基金资助项目(JC20120239)

摘　　要：针对传统的CPU人脸检测程序难以满足高清图像实时检测要求的问题,本文提出一种在CUDA平台下基于Viola-Jones级联分类器的快速人脸检测算法,其中积分图计算、扫描窗口检测和分类器放大修正部分均进行了并行加速。扫描窗口检测采用的窗口重映射技术可有效解决非规则树结构应用的负载不均衡问题。实验结果表明,对于1080p图像,CUDA程序与CPU和OpenCV程序相比可分别实现17.04、3.22倍的加速比,同时具有相近的检测精度。For the traditional CPU facial detection program is difficult to meet the requirements of real-time detection over high definition images, this paper proposes a fast face detection algorithm based on the Viola-Jones cascade classi/ier in the CUDA platform. It implements and improves novel parallel methodologies of image integral calculation, scan window processing and the amplification and correction of classifiers. The load imbalance problem of irregular tree structure applications could be effectively solved by the window re-mapping technology used in the scan window processing. The experimental results show that the CUDA program could respectively achieve 17.04, 3.22 times speedup compared with the CPU and OpenCV programs for the 1080p im- age, while maintaining a similar detection quality.

关 键 词：人脸检测 级联分类器 积分图 计算统一设备架构 HAAR-LIKE特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]