基于注意力的热点块和显著像素卷积神经网络的人脸防伪方法

人脸防伪用于验证被测试者是否为真实活体,是计算机视觉领域的一个研究热点。攻击手段的多样性以及人脸识别主要在嵌入式、移动式等不具备高计算能力的设备上应用,使得快速有效的人脸防伪计算成为具有挑战性的任务。针对该问题,文中提出了一种基于注意力的热点块和显著像素卷积神经网络的方法。其中,热点块机制以对5个热点块的判别来取代对整张人脸的判别,显著降低了计算量,迫使网络模型集中关注更具有鉴别信息的热点块,提高了网络模型的准确率;显著像素方法对输入的人脸图像进行显著像素预测,通过判断显著预测图是否符合人脸的深度特性来鉴别活体与攻击。该方法将热点块与显著像素的结果进行融合,充分发挥了局部特征和全局特征的作用,进一步提升了人脸防伪的效果。与现有方法相比,所提方法在CASIA-MFSD、Replay-Attack以及SiW数据集上都达到了很好的效果。

Attention-based Hot Block and Saliency Pixel Convolutional Neural Network Method for Face Anti-spoofing

Face anti-spoofing is used to verify whether the testee is a real person.The diversity of attack methods and the application of face recognition on various embedded and mobile devices with low computing capabilities have made face anti-spoofing a very challenging task.Aiming at face anti-spoofing,an attention-based hot block and saliency pixel convolutional neural network method is proposed.The hot block method replaces the discrimination of the entire face with the determination of 5 hot blocks,which not only reduces the amount of calculation,but also forces the network to focus on hot spots with more discerning information,so as to improve the accuracy of the network.On the other hand,the saliency pixel method performs saliency pixel prediction on the input face image to determine whether the saliency prediction map meets depth characteristics of the face to identify the liveness and the attack.This method fuses the results of hot blocks and saliency pixels to give full play to the role of local features and global features,and further enhances the effect of face anti-spoofing.Compared with existing methods,the proposed method has achieved good results on CASIA-MFSD,Replay-Attack and SiW datasets.