Attentive Contrast Learning Network for Fine-Grained Classification

Fine-grained visual classification is challenging due to subtle differences between sub-categories. Current popular methods usually leverage a single image and are designed by two main perspectives: feature representation learning and discriminative parts localization, while a few methods utilize pairwise images as input. However, it is difficult to learn representations discriminatively both across the images and across the categories, as well as to guarantee for accurate location of discriminative parts. In this paper, different from the existing methods, we argue to solve these difficulties from the perspective of contrastive learning and propose a novel Attentive Contrast Learning Network (ACLN). The network aims to attract the representation of positive pairs, which are from the same category, and repulse the representation of negative pairs, which are from different categories. A contrastive learning module, equipped with two contrastive losses, is proposed to achieve this. Specifically, the attention maps, generated by the attention generator, are bounded with the original CNN feature as positive pair, while the attention maps of different images form the negative pairs. Besides, the final classification results are obtained by a synergic learning module, utilizing both the original feature and the attention maps. Comprehensive experiments are conducted on four benchmark datasets, on which our ACLN outperforms all the existing SOTA approaches. For reproducible scientific research https://github.com/mpskex/AttentiveContrastiveLearningNetwork.