A Nonuniform Weighted Loss Function for Imbalanced Image Classification
Qiang Qiu, Zichen Song · 2018
In the field of deep image classification, it is a challenging task to learn the classifier from imbalanced dataset. Existing deep neural networks usually equally punish the training loss of each sample. To minimize the average loss, the deep classifier inevitably tends to sacrifice the classification accuracy of the minority when the sizes of training samples are imbalanced across different categories. To address this issue, this paper proposes an nonuniform weighted loss function which aims to compensate the bias of training loss for the minority categories. The recent popular datasets are balanced in terms of the sample size across different classes. So we build a new large scale imbalanced dataset to verify the proposed method. Our method start from deriving the gradient formulation of the proposed loss function, and a more complete explanation is given to demonstrate its effectiveness.