A Two -Stage Active Learning Method for Image Classification

Recent successes in learning -based image classification rely heavily on a large number of annotated training samples, which often require considerable human effort. In this paper, we propose a new two -stage active learning (AL) method for image classification with a query strategy considering both uncertainty and diversity. In the first stage, the uncertainty is used to determine the candidate set. In the second stage, the candidate set is clustered, and the nearest sample from the cluster center is selected to increase the diversity of samples. Our method uses Poly -I loss as the classification loss and Binary Cross Entropy (BCE) as the binary loss to distinguish between the labeled samples and the unlabeled samples. We train the two classifiers in a joint way and evaluate our method on the MARIO and CIPAR100 datasets. The rich experimental results show that the proposed method outperforms the state-ofthe-art AL methods in image classification tasks.