二代证人脸识别的多姿态虚拟样本生成方法

Multi-pose virtual sample generation for face recognition using the second generation ID card

使用第二代身份证照片作为训练样本进行人脸识别属于典型的单样本问题,由于没有充分数量的训练样本,会造成常规的人脸识别算法识别率低下,甚至无效的问题。为此采用虚拟样本生成方法,并针对遇到姿态变化较复杂的人脸时,识别率不高的问题,提出了一种新的多姿态的虚拟样本生成方法,通过模拟人脸侧向旋转、俯仰和立体旋转等增加有效的训练样本,再使用鲁棒性较好的HMM进行人脸识别。在自建的身份证人脸库上进行测试,实验结果显示,该方法在一定程度上减弱了人脸姿态的变化对识别率的影响,并取得了较好的识别效果。

Using the Chinese second generation identity card＇s （2G-ID card） photo as the training sample for human face recognition is a typical single-sample problem. The insufficiency of training samples will cause the serious performance reduction of conventional face recognition algorithms. Virtual sample generation is an approach to transfer single sample to multi-sample thus re-work many face recognition methods. However, Existing virtual sample generation technology cannot efficiently deal with faces with diverse pose. A new multi-pose virtual sample generation technology is proposed. Through the face simulation of lateral rotation, pitching and three-dimensional pose variation of human＇s head to increase the training samples, then using HMM to realize face recognition, The experimental results show that the method can overcome the affect of changes in face pose to some extent, and can obtain higher recognition performance.