基于卷积神经网络(CNN)和CUDA加速的实时视频人脸识别

Real-time Face Recognition in Videos Based on Convolutional Neural Networks(CNN) and CUDA

为了兼顾视频人脸识别中识别准确率和实时性,提出了基于卷积神经网络(CNN)和CUDA加速的实时视频人脸识别方法。构建了一个6层结构的CNN人脸识别网络,在视频帧中通过Adaboost算法检测到的人脸输入所构建的CNN中进行视频人脸识别,结合CUDA并行计算架构,对算法进行加速。此外为了更适用于实际视频监控情况,通过对CNN网络结构末尾Softmax分类器的分类结果进行多级判决引入了开集人脸识别功能。从多个角度对该方法进行了实验验证,结果证明,此方法可满足识别准确率和实时性要求,同时对于视频中人脸姿态变化、光照变化、距离远近等都具有良好的鲁棒性。

Aiming at the recognition rate and time-consuming of face recognition in videos,a real-time videobased face recognition method is proposed based on convolutional neural networks（ CNN） and CUDA. A 6-layer CNN was built,and the faces in video frames detected by Haar Adaboost will be entered into the CNN. The whole process was accelerated by CUDA. In addition to be more suitable for the actual situation,open-set face recognition was introduced by multistage decision which process the results of the Softmax classifier. Experimental results show that the recognition rate is high,while the time-consuming is satisfactory. Otherwise,this method has high robustness of the face pose-change,illumination-change and the distance in videos.