基于深度学习的人脸识别方法研究进展

Research Progress of Face Recognition Based on Deep Learning

基于深度学习的人脸识别方法主要分为两个方向:多分类和度量学习。多分类的方法在标记的已知类别上训练,在未知类别上测试。测试集上的识别性能严重受限于训练集上模型的表达能力,近几年的研究工作主要是基于分类损失函数的改进,动机在于让模型在闭集上学习的特征具有更高的辨别性。度量学习的动机在于学习新的表征使得类间距离大于类内距离,训练阶段不需要知道目标的具体类别,只需要标记类别差异。近几年对于度量学习方向的研究工作主要集中在损失函数的改进,调整不同的策略减小类内距离方差,同时增大类间距离方差,学习到的度量可以直接作为特征比对阶段的相似度。对这两个方向的研究工作进行归纳和总结,并对其他可能的方向做一些展望,为基于深度学习的人脸识别方法的进一步研究提供一些参考。

Face recognition methods based on deep learning are mainly divided into two directions:multi classification and metric learning.The multi classification method is trained on the known classes of tags and tested on the unknown classes.In recent years,the research work is main⁃ly based on the improvement of the classification loss function.The motivation is to make the characteristics of the model learning on the closed set more discriminative.The motivation of metric learning is to learn new representation so that the distance between classes is great⁃er than the distance within the class.In the training stage,we do not need to know the specific category of the target,only need to mark the category difference.In recent years,the research on metric learning mainly focuses on the improvement of loss function,adjusting different strategies to reduce the same class distance variance,and increasing the different class distance variance.The learned metric can be direct⁃ly used as the similarity of feature comparison stage.Summarizes the research work of these two directions,and makes some prospects for other possible directions,which provides some reference for the further research of face recognition method based on deep learning.