Hard-Mining Loss based Convolutional Neural Network for Face Recognition

TL;DR

This paper introduces a Hard-Mining loss function that emphasizes hard samples during training, improving face recognition accuracy when combined with existing loss functions like ArcFace on standard datasets.

Contribution

It proposes a generic Hard-Mining loss that enhances existing loss functions by prioritizing hard samples, leading to improved face recognition performance.

Findings

Hard-Mining loss boosts accuracy of existing loss functions.

Improved performance on LFW and YTF datasets.

Effective with multiple loss functions like ArcFace and Softmax.

Abstract

Face Recognition is one of the prominent problems in the computer vision domain. Witnessing advances in deep learning, significant work has been observed in face recognition, which touched upon various parts of the recognition framework like Convolutional Neural Network (CNN), Layers, Loss functions, etc. Various loss functions such as Cross-Entropy, Angular-Softmax and ArcFace have been introduced to learn the weights of network for face recognition. However, these loss functions do not give high priority to the hard samples as compared to the easy samples. Moreover, their learning process is biased due to a number of easy examples compared to hard examples. In this paper, we address this issue by considering hard examples with more priority. In order to do so, We propose a Hard-Mining loss by increasing the loss for harder examples and decreasing the loss for easy examples. The proposed concept is generic and can be used with any existing loss function. We test the Hard-Mining loss with different losses such as Cross-Entropy, Angular-Softmax and ArcFace. The proposed Hard-Mining loss is tested over widely used Labeled Faces in the Wild (LFW) and YouTube Faces (YTF) datasets. The training is performed over CASIA-WebFace and MS-Celeb-1M datasets. We use the residual network (i.e., ResNet18) for the experimental analysis. The experimental results suggest that the performance of existing loss functions is boosted when used in the framework of the proposed Hard-Mining loss.