改进型MobileNetV3轻量级人脸活体检测算法

LIGHTEWIGHT FACE ANTI-SPOOFING ALGORITHM BASED ON IMPROVED MOBILENETV3

为了解决人脸识别系统中的欺诈攻击问题,以及目前基于深度学习的活体检测方法大多以大型卷积网络作为主干网络,导致模型结构复杂、计算量大等问题,提出一种改进型MobileNetV3的轻量级人脸活体检测算法。对MobileNetV3中利用全局平均池化计算通道注意力权重和使用双非线性全连接层存在的不足进行讨论,提出新的注意力机制EFCANet,并利用EFCANet网络对MobileNetV3轻量级卷积神经网络进行改进。实验结果表明,改进后轻量型活体检测算法,在检测精度、网络模型大小、损失值和等错误率等方面有着不错的表现。

In order to solve the problem of fraud attacks in face recognition systems and the problem that most deep learning based in-vivo detection methods use large convolutional networks as the backbone network,which leads to complex model structure and large computation amount,an improved lightweight face detection algorithm based on MobileNetV3 is proposed.This paper discussed the shortcomings of using global average pooling to calculate channel attention weight in MobileNetV3 and using double nonlinear fully connected layer.A new attention mechanism EFCANet was proposed,and the EFCANet network was used to improve MobileNetV3 lightweight convolutional neural network.The experimental results show that the improved light-weight face anti-spoofing algorithm has a good performance in detection accuracy,network model size,loss value,and equal error rate.