摘要
针对视频人脸识别中存在的动态人脸信息捕捉困难和局部人脸特征提取粗糙的问题,提出了一种基于深度Q学习和注意模型结合的视频人脸识别方法。首先,采用卷积神经网络(Convolutional Neural Network,CNN)训练视频数据可提取多维特征;其次,将视频特征输入注意模型,根据视频数据时间连续性信息得到局部人脸特征、人脸位置和时间记忆单元;最后,采用Q学习迭代计算注意模型的输出,找到含人脸的最优帧序列,并以此计算视频匹配准确度。实验结果表明,该方法有效提高了复杂背景下视频人脸识别的准确性。
Aiming at the difficulty of capturing dynamic face information and the rough extraction of local facial features in video face recognition,a video face recognition method based on deep Q learning and attention model is proposed.Firstly,the Convolutional Neural Network(CNN)training video data can extract multi-dimensional features.Then,the video features can input into the attention model,and the local face features,the face positions and the time memories unit are obtained according to the temporal continuity information of the video data.Then,the Q learning is used to calculate the output of the attention model,and the optimal frame sequence containing the face is found,and calculates the video matching accuracy.The experimental results show that the method effectively improves the accuracy of video face recognition in complex backgrounds.
作者
郑秋文
刘惠义
ZHENG Qiu-wen;LIU Hui-yi(School of Computer and Information,Hohai University,Nanjing 211100,China)
出处
《信息技术》
2019年第4期111-115,120,共6页
Information Technology
