单样本多姿态人脸识别研究

Various pose face recognition with one front training sample

对如何用单幅正面人脸图像进行训练,待识别图像具有多种姿态变化的人脸识别问题进行了研究。人脸识别算法的识别率常与每人的训练样本数正相关。但在实际应用中,要求每人提供多幅图像并不合理。通过增加虚拟图像提高识别率,给出了一种模拟人脸姿态改变后的近似图像的简单有效的算法。在FERET人脸库上的实验表明,该文提出的近似图像对提高识别率作用显著,最好识别率提高了28·2%。

Almost all algorithms for face recognition have tight relationship with the images number of each person. The recognition rate increases with the increasing training number of each class. But in applications, it is not practical to ask for many training images from each person. A new method, which can generate the simulated images of face after rotating an angle, was proposed. It generalized the method of Fisherfaces and uncorrelated image projection diseriminant analysis to one sample per person. The recognition rates of Principal Component Analysis （ PCA）, Fisherfaces, and Two dimension＇s PCA （2DPCA） were also studied. The experimental results on FERET face-databases indicate that after adding virtual images, the recognition rates increase greatly, and the best recognition rate has improved by 28.2%.