CUDA-based real-time face recognition system

This paper proposes a real-time face recognition system based on the Compute Unified Device Architecture (CUDA) platform, which effectively completed the face detection and recognition tasks. In the face detection phase with Viola-Jones cascade classifier, we implemented and improved novel parallel methodologies of image integral, calculation scan window processing and the amplification and correction of classifiers. In the face recognition phase, we explored the parallelizing of the algorithm and parallelized some part of the testing phase. Through the optimization of the two important part of face recognition system, the system we proposed make a big difference. The experimental results demonstrate that, in comparison with traditional CPU program, the proposed approach running on an NVidia GTX 570 graphics card could respectively achieve 22.42 times speedup in detection phase and 1668.56 times speedup in recognition phase when only training 2000 images and testing 40 images compared with the CPU program running on an Intel core i7 processor. The recognition speed will increase until it reaches the hardware resource limit. It shows that the system we proposed achieves a good real-time performance.