基于多任务CNN的人脸活体多属性检测

人脸的活体检测是人脸识别系统的安全性保证,传统交互式活体状态识别中通常先进行关键点检测,再进行人脸活体状态判断,无法利用活体属性之间关联性同时进行多种活体属性检测。基于多任务卷积神经网络(convolutional neural network,CNN),提出了一种直接从输入人脸中同时判断人脸的眼睛睁闭状态、嘴巴的张闭状态、头部点头状态和摇头中左右侧脸状态四个任务的活体检测方法。该方法利用多层CNN强大的特征提取能力、多任务的并行能力和四个任务中的关联性,直接提取人脸照片中的特征信息,判断人脸多个活体属性。实际应用结果表明,这种基于多任务CNN方法的准确率在四个任务中均可达95%以上,甚至达到98%以上,无论检测的准确率还是同时检测多个任务的能力均明显优于传统的人脸活体属性检测方法。

Face Liveness Multi-Attribute Detection Via Multi-Task CNN

Liveness detection is the security guarantee for face recognition systems. Traditional interactive liveness detection methods usually detect liveness based on predetermined facial feature points. They could not detect multiple liveness attributes simultaneously according to the relevance of different face liveness attributes. In this paper, a new liveness detection method which can concurrently detect four liveness attribute states of input face, eye state, mouth state, nod head state and shake head state, is proposed based on multi-task CNN(Convolutional Neural Network). This method takes advantage of the relevance of the four face attributes and the powerful feature extraction and parallel multi-task abilities of CNN to detect multiple face liveness attributes with feature information extracted directly from face image. It is shown by experimental results that the detection accuracy rate of proposed method for all of the four attributes can reach 95% and sometimes even more than 98%. Both the accuracy rate and multi-task ability of proposed method are superior to traditional interactive face liveness detection methods.