Abstract: In recent years, the sample mining strategy has been integrated into the loss function of face recognition, significantly improving the performance of face recognition. But most of the work focuses on how to mine difficult samples during the training phase, without considering the potential unrecognized sample images in the difficult samples, resulting in poor recognition performance of the model for low-quality facial images. To solve this problem, this paper proposes a hybrid adaptive loss function MixFace that combines sample difficulty adaptation and image quality adaptation. The loss function combines the CurricularFace based on curriculum learning with the image adaptive loss function AdaFace. The feature norm is incorporated into the loss function as an image quality indicator. On the premise of focusing on image quality, this paper focuses on simple samples in the early training stage and difficult samples in the later training stage, reducing the network model’s attention to some low-quality unrecognized samples in difficult samples. Trained on CASIA-WebFace and MS1MV2 datasets, MixFace shows varying degrees of performance improvement compared with CurricularFace and AdaFace on high-quality test sets LFW, CFP_FP, AgeDB, CALFW, and CPLFW. At the same time, MixFace shows better recognition performance than CurricularFace and AdaFace on medium quality test sets IJB-B, IJB-C and low-quality test set TinyFace. Experimental results show that MixFace can effectively reduce the interference of unrecognized images, thereby improving the performance of low-quality face recognition. At the same time, benefiting from the curriculum learning method in MixFace, it can still maintain good performance for high-quality face recognition.

Key words: face recognition, curriculum learning, image quality, adaptive loss

摘要： 近年来样本挖掘策略被融入人脸识别的损失函数中，显著提升了人脸识别性能，但大部分工作都集中在如何在训练阶段挖掘困难样本，没有考虑到困难样本中潜在的无法识别的样本图像，从而导致模型对低质量人脸图像的识别性能较差。针对该问题，提出了一种结合样本难度自适应和图像质量自适应的混合自适应损失函数MixFace。该损失函数将基于课程式学习的损失函数CurricularFace与图像自适应损失函数AdaFace相结合，将特征范数作为图像质量指标融入损失函数中，在关注图像质量的前提下在训练前期关注简单样本，后期关注困难样本，降低网络模型对困难样本中部分低质量不可识别样本的关注。分别使用CASIA-WebFace和MS1MV2数据集训练，MixFace在高质量测试集LFW、CFP_FP、AgeDB、CALFW和CPLFW上相比Curricular-Face和AdaFace有不同程度的性能提升，同时MixFace在中等质量测试集IJB-B、IJB-C以及低质量测试集TinyFace上显示出比CurricularFace和AdaFace更好的识别性能。实验结果表明，MixFace能有效降低无法识别图像的干扰，进而提升低质量人脸识别性能，同时受益于MixFace中课程式学习的方式，对于高质量人脸识别仍然能保持较好的性能。

关键词: 人脸识别, 课程式学习, 图像质量, 自适应损失