基于OpenCL的Viola-Jones人脸检测算法性能优化研究

发布时间：2018-09-12 17:05

【摘要】：Viola-Jones人脸检测算法是最为成功的可实用的人脸检测算法之一.然而,随着该算法所在领域数据处理规模的不断扩大,现有算法的性能已经越来越无法满足日益增长的交互性与实时性要求.使用GPU计算平台提升该算法性能,以满足日益增长的实时性要求已经成为研究热点.然而,该算法在对GPU的实现和优化中,存在线程间负载不均衡的非规则特性,如果仅使用传统的优化方法,则难以在GPU计算平台上达到较高性能.针对此种情况,该文构建了针对此类算法的并行优化框架,通过Uberkernel、粗粒度并行、Persistent Thread、线程与数据的动态映射、全局及本地队列等优化方法的应用,突破了负载不均衡非规则特性导致的性能瓶颈,大幅提高了人脸检测算法在GPU计算平台上的性能.同时,该文通过对不同GPU计算平台关键性能参数的定义、抽取和传递,实现了该算法在不同GPU计算平台间的性能移植.实验结果表明,与OpenCV2.4中经过高度优化的CPU版本在Intel Xeon X5550CPU上的性能相比,优化后的算法在AMD HD7970和NVIDIA GTX680两个不同GPU计算平台上分别达到了11.24~20.27和9.24~17.62倍的加速比,不仅实现了高性能,而且实现了在不同GPU计算平台间的性能移植.
[Abstract]:Viola-Jones face detection algorithm is one of the most successful and practical face detection algorithms. However, with the expansion of data processing scale in the field of the algorithm, the performance of existing algorithms has become more and more unable to meet the increasing requirements of interactivity and real-time. Using GPU computing platform to improve the performance of the algorithm to meet the growing real-time requirements has become a research hotspot. However, in the implementation and optimization of GPU, the algorithm has the irregular property of load imbalance between threads. If only traditional optimization methods are used, it is difficult to achieve high performance on GPU computing platform. In this paper, a parallel optimization framework for this kind of algorithm is constructed, and the application of optimization methods such as dynamic mapping of Uberkernel, coarse-grained parallel persistent Thread, threads and data, global and local queues, etc. It breaks through the performance bottleneck caused by the irregular characteristics of load imbalance and greatly improves the performance of face detection algorithm on GPU computing platform. At the same time, by defining, extracting and transferring the key performance parameters of different GPU computing platforms, this paper realizes the performance migration between different GPU computing platforms. The experimental results show that compared with the highly optimized CPU version of OpenCV2.4 on Intel Xeon X5550CPU, the optimized algorithm achieves a speedup ratio of 11.24% 20.27 and 9.24N 17.62 times on two different GPU platforms, AMD HD7970 and NVIDIA GTX680, respectively. The proposed algorithm not only achieves high performance, but also achieves high performance. Moreover, the performance migration between different GPU computing platforms is realized.
【作者单位】：中国科学院计算技术研究所计算机体系结构国家重点实验室;
【基金】：国家自然科学基金(61133005,61272136,61521092,61402441)资助~~
【分类号】：TP391.41

【相似文献】