The National Natural Science Foundation of China (61771085); The Project of Key Laboratory of Signal and Information Processing of Chongqing (CSTC2009CA2003)
针对人脸信息认证系统中存在的欺骗攻击问题，利用色彩空间转换信息丢失的特性，提出一种人脸活体检测算法。通过Gabor滤波器组多尺度、多方向地增强关键人脸图像纹理特征，抑制人脸图像的一般特征。高斯径向基函数分类器分类SURF(speeded up robust features)算子提取特征描述子，区分人脸活体与非法用户的欺骗攻击。利用类间方差衡量特征改进前后的可分性，计算图像原始特征与纹理增强后特征的类间方差、类内方差大小以及可分性判据J。在公开数据集Replay-Attack,CASIA-FASD数据库进行测试，彩色纹理图像增强后，人脸关键特征被增强而一般特征被抑制，背景与目标的类间方差增大，类内方差减小，特征可分性增强从而更具鲁棒性，使得纹理细节能被有效利用。实验结果表明，该算法能有效、实时地判断人脸活体与欺骗攻击。
Aiming at the problem of spoofing attack in face information authentication system, this thesis proposes a face detection algorithm by using information loss characteristic in color space transformation. Gabor filter banks are used to enhance the texture features of face images in multi-scale and multi-direction. Gauss Radial Basis Function Classifier classifies the feature descriptors extracted by SURF operator, and distinguishes face living objects and the spoofing attacks of illegal users. The inter-class variance, intra-class variance and separability criterion J of the original image features and texture enhanced features are calculated by using the inter-class variance to measure the separability before and after the feature improvement. Finally, the algorithm is tested in public data dataset Replay-Attack and CASIA-FASD. After the enhancement of the color texture image, the key features of the face are enhanced and the general features are suppressed. The variance between the background and the target is increased, and the intra-class variance is reduced. Feature separability is enhanced to be more robust,allowing texture details to be effectively utilized.. The experimental results show that the proposed algorithm can effectively judge the human face and spoofing attacks in real time.