基于多模态融合的活体检测研究

Face Anti-spoofing Based on Multimodal Fusion

活体检测技术是人脸识别系统中的一个重要环节,基于深度学习的方法依赖RGB模态,难以提取更有效的图像特征,因此提出了基于多模态特征融合的活体检测方法。首先,提取RGB、深度和红外3个模态在卷积神经网络中的高层语义特征,并进行融合。其次,为了提取更具辨别性的特征信息,将网络的输入改为图像的局部区域块。最后,在模型训练的过程中加入多模态特征随机擦除操作,可有效降低过拟合,提升模型鲁棒性。实验表明:融合多模态特征能够有效识别不同类型的假体攻击,为人脸识别提供安全保障。

Face anti-spoofing is an important technology in the face recognition system.The method based on deep learning has achieved good results in this field.However,the previous methods only focus on RGB modal feature,and it is difficult to extract richer feature information.Therefore,we propose a face anti-spoofing method based on multi-modal feature fusion.Firstly,we extract the high-level semantic features of the three modalities of RGB,depth and infrared in the convolutional neural network,and concatenate them.Secondly,in order to extract more discriminative feature information,we change the input of the network to a local patch the face image.Finally,the random erasing operation of multi-modal features is added in the process of model training,which can effectively reduce over-fitting and improve the robustness of the model.Experiments show that the fusion of multi-modal features can effectively identify different types of attacks,and provide security for face recognition.