Improved combined margin loss function for face recognition

Abstract

In the face recognition field, classification through softmax is difficult, and a marginbased loss function that considers the angle between the feature vector and the class center shows good performance. Therefore, this work focuses on the study of loss functions rather than building deep convolutional neural networks. This work explores the Softmax loss function from the properties of the MagFace loss function, which has recently shown good performance. This work proposes a new loss function that combines these properties called Improved Combined Margin. The new loss function considers the need for variable margin in angular space and considers margin in angular space and cosine space, simultaneously. Compared with the existing loss functions, experiments are conducted on various face recognition datasets, and the loss function proposed in this work increases the face recognition performance.