Focal Loss Improves the Model Performance on Multi-Label Image Classifications with Imbalanced Data

Abstract

Dealing with imbalanced data has always been a great challenge in statistical learning. In classification problems, instances for different classes are unequal in an imbalanced dataset. Therefore, traditional machine learning classification algorithms are usually sensitive to this imbalance both in the training and inference processes. In this paper, by addressing the class imbalance on the basis of Focal Loss, we introduce an approach to improve the performance of convolutional neural networks (CNNs) on the multi-label image classification with an extremely imbalanced dataset. As focal loss puts more focus on hard and misclassified examples when comparing with classic cross-entropy loss, our results demonstrate that such loss function indeed achieves a significant improvement of CNN models (ResNet-50, ResNet-101, and SE-ResNeXt-101) regarding the mean F2 scores on the test set.