EnsFace: An Ensemble Method of Deep Convolutional Neural Networks with Novel Effective Loss Functions for Face Recognition

Abstract

In recent years, designing new effective Deep Convolutional Neural Network (DCNN) architectures and loss functions are two crucial trends in improving face recognition (FR) accuracy. However, building an optimal DCNN and ameliorating FR performance are still the main challenges for researchers. Thus, we first investigate and analyzes the effect of several novel effective loss functions based on softmax on DCNN with the ResNet architecture. We then propose an ensemble learning, namely EnsFace, by taking advantage of recent novel FR methods based on CosFace, ArcFace, and MagFace. EnsFace elaborates on the voting mechanism that utilizes non-optimal pre-trained models to obtain better discriminative ability for FR. To prove the performance of both speed and accuracy of EnsFace, we carry out rigorous experiments using several popular benchmarks, including LFW, CFP-FP, and AgeDB-30, as well as two renovations of LFW: CALFW and CPLFW. The results of our experiments achieve state-of-the-art figures, which show the proposed method’s massive potential in improving FR performance. Ablation studies and overall benchmarks undeniably prove the effectiveness of our EnsFace.

Recommended citation:

update late

Download paper here