Deep active learning models for imbalanced image classification

Abstract

Active learning can query valuable samples in an unlabeled sample pool for annotation, thus building a more informative labeled dataset and reducing the annotation cost. However, traditional active learning methods are not effective in the task of imbalanced image classification for ignoring the distribution bias. In this study, we propose a Balanced Active Learning (BAL) method for imbalanced image classification. BAL estimates the probability of a sample belonging to minority or majority classes and compensates the annotation query for the minority classes, thus alleviating the class imbalance in the selected training samples. Experiments on three imbalanced image classification datasets, imbalanced CIFAR-10, ISIC2020, and Caltech256, showed that BAL achieved new state-of-the-art performance of active learning in a variety of classification tasks and different types of imbalance.