Application of Difficult Sample Mining based on Cosine Loss in Face Recognition

Due to the development of deep convolutional neural networks, face recognition has made great progress, and its main goal is how to improve feature recognition capabilities. In this regard, several loss functions based on angular boundaries have been proposed to increase the feature margin between different classes. Although very good results have been achieved in this direction, there are still some problems. These loss functions only expand the feature margin from the perspective of real classification in training, and do not provide distinguishability for misclassified samples. In order to solve this problem, this paper improves on the original cosine loss function, and implements feature learning in the direction of difficult samples based on misclassified feature vectors.