Dynamic Attention Loss for Small-Sample Image Classification

Convolutional Neural Networks (CNNs) have been successfully used in various image classification tasks and gradually become one of the most powerful machine learning approaches. To improve the capability of model generalization and performance on small-sample image classification, a new trend is to learn discriminative features via CNNs. The idea of this paper is to decrease the confusion between categories to extract discriminative features and enlarge inter-class variance, especially for classes which have indistinguishable features. In this paper, we propose a loss function termed as Dynamic Attention Loss (DAL), which introduces confusion rate-weighted soft label (target) as the controller of similarity measurement between categories, dynamically giving corresponding attention to samples especially for those classified wrongly during the training process. Experimental results demonstrate that compared with Cross-Entropy Loss and Focal Loss, the proposed DAL achieved a better performance on the LabelMe dataset and the Caltech101 dataset.