SWS-DAN: Subtler WS-DAN for fine-grained image classification

摘要

Currently, weakly supervised data augmentation network (WS-DAN) has been proved to be one of the state-ofthe-art methods for fine-grained image classification due to its effectiveness on attention-guided data augmentation and bilinear attention pooling. Taking WS-DAN as the backbone, in this paper, we further propose a subtler WS-DAN recognition network, namely, SWS-DAN. Specifically, we first construct a novel "salience-guided data augmentation" scheme composed of cutblock, part-aware cropping, and SCutMix operations, which can more effectively expand the number of training dataset and improve the weakness addressed in the data augmentation procedure of WS-DAN. Meanwhile, the novel data-augmentation manner reduces background noise and mines more discriminative regions simultaneously, thereby avoiding the overfitting. In caring about the key issue in fine-grained image classification task is how to distinguish the extremely similar subclasses (e.g., Artic Tern, Elegant Tern, and Forsters Tern), we then design a "Top-k" loss function that mainly focuses on the similar classes so as to find their extraordinary subtle differences. Extensive experiments carried out on common fine-grained image datasets demonstrate that SWS-DAN can surpass WS-DAN with a significant margin in the classification performance.