Norm-Integrated Softmax Loss For Deep Face Recognition

The key to face recognition lies in how to improve the model’s ability to extract facial features. To this end, numerous loss functions based on different metrics have been proposed to increase the margin of feature distinction between different classes. Methods based on Cosine distance significantly enhance face recognition performance by focusing on angular constraints between samples and classes, demonstrating their superiority over those based on Euclidean distance. However, a significant oversight in these methods is the neglect of feature magnitude’s importance in representing facial features. To address this gap, our study introduces the NormIntegrated Softmax loss (NIface loss), a novel loss function that amalgamates feature norms with angular information. This integration offers a comprehensive perspective for feature classification, augmenting the compactness of intra-class features. Extensive evaluations on large-scale public datasets have demonstrated the efficacy of NIface loss in enhancing recognition accuracy and stability.