ICNN: An Inter-class Nearest Neighbor Method for Face Recognition
Hongxia Zhang, Tao Wang, Erzhong Gao, Minqiang Xu, Liang Ju He, Jianshe Wang, Lin Liu · 2023
Face recognition has achieved rapid development in recent years. An appropriate loss function is a key part of the model to learn distinguished features. The following facts may be considered: 1) the margin-based softmax methods fail to pay attention to those hard/semi-hard samples, while model can learn discriminating features from those samples. 2) the mining-based softmax methods define the boundary between easy and hard samples relatively strictly. Based on this, we propose a novel "Inter-class Nearest Neighbor" (ICNN) loss function, that selecting k samples not belonging to yi-th class and pushing those samples away from the prototype, to enlarge inter-class distance and enable model to learn more discriminating features. Our method can obtain orthogonal improvements to existing margin-based softmax loss. Extensive experiments demonstrate that proposed ICNN method can achieve competitive performance on the popular test datasets, especially outperform prior methods on more challenging IJB datasets.