TIO Loss: A Transplantable Inversed One-Hot Loss for Imbalanced Multi-classification

摘要

In image classification, class imbalance is a common problem when training neural networks. It is partly because collecting a great quantity of images in reality is a difficult task. Class imbalanced datasets usually lead to imbalanced learning, especially in multi-classification. In this paper, we introduce our inversed one-hot learning method by invers-ing the encoding way of labels and present a transplantable inversed one-hot loss, named TIO loss, which can be added to current existing loss functions. Our main idea is using the penalty property of logarithm function by taking the misclassified classes into consideration. We conduct the ablation experiments by adding TIO loss to cross-entropy loss and focal loss. Finally, we verify our method on two imbalanced datasets and the experimental results show the significant improvements.