基于姿态校正与虚拟样本的多姿态人脸识别

Pose-varied Face Recognition Based on Facial Pose Correction and Virtual Samples

研究人脸识别问题,提高多姿态识别精度。针对训练样本不足,当测试人脸图像姿态变化较大时,就会降低一致性,使得识别精度急剧下滑(低于60%),甚至出现无法识别的情况。为解决因训练样本不足导致识别精度低下的问题,根据正弦变换的改进型姿态校正人脸识别策略,在保留人脸图像的纹理信息的情况下,将多姿态样本校正为正面人脸图像,利用二次多项式变换方法增加虚拟训练样本,解决了实际情况中只能获取一个正面或侧面训练样本的问题,于是采用子空间的特征提取方法进行仿真,在保证时间消耗的情况下,识别率相比传统模型提高了19个百分点,达到77%,表明改进方法能对多姿态人脸进行有效识别,并提高了识别精度。

Improving the multi-gesture recognition accuracy in the study of face recognition.The recognition rate will decline sharply when there are large variation of face pose,especially when the training samples are small,the identification may be impossible.In order to solve the problem of lack of training samples,firstly this paper transforms the pose-varied face images to frontal face images and the texture information of faces is kept based on sine transform（ST）.Secondly,polynomial transform is used to generate virtual samples when only having single training sample.Finally,combining the subspace feature extraction methods with pose-varied face recognition strategy,the simulation is conducted and the recognition rates have increased by 19 percentage points.