一种增强人脸识别模型训练稳定性的损失函数

随着卷积神经网路的快速发展，深度学习在人脸识别领域进行了大量的应用。近几年，人脸识别准确率快速提高，主要归功于新颖损失函数的提出。在目前最大的人脸评测数据集MegaFace上，最顶尖的模型已经实现了97.91%的1∶N查找性能，但是训练过程中收敛稳定性问题没有得到解决。该文提出一种新型的损失函数LineFace，其logit曲线在余弦空间中呈线性，使训练中梯度收敛更加稳定。大量实验表明，该损失函数可以实现良好的模型性能收敛性与识别性能。

New Loss Function to Enhance the Training Stability of Face Recognition Model

With the rapid development of convolutional neural networks, deep learning has been widely used in the field of face recognition. In recent years, the accuracy of face recognition has increased rapidly, mainly due to the proposition of novel loss functions. On the current largest face test set MegaFace, the top model has achieved 97.91% 1∶N search performance, but the problem of convergence stability during training has not been properly solved. Thus a new type of loss function, Lineface, is proposed in this paper. Its logic curve is linear in the cosine space, which makes the gradient convergence better and more stable during training. A large number of experiments show that good model performance and convergence can be achieved.