Learn more: Sub-significant area learning for fine-grained visual classification

Fine-grained visual classification is more challenging as a subtask of image classification due to the large intra-class and slight inter-class variations. Recent work has focused on localizing discriminative features using attentional mechanisms. However, attention tends to focus on the salient parts of feature maps, ignoring other regions that are not salient but are discriminative for fine-grained classification. In this regard, we propose an efficient method called Discriminative Region Learning Dual-Branch Attention Network (DAL-Net) to address this problem. We propose: (1)Avoid over-focusing on local features by pixel-level attention wipe on salient features. (2)The features are enhanced and suppressed by the channel space enhancement module to mine multiple discriminative regions. (3)To learn complementary semantic information, we fuse cross-regional discriminative features from another branch. Our method can be trained end-to-end without bounding boxes and annotations. We demonstrate that our method can obtain competitive results on the CUB200-2011, FGVC-Aircraft, Stanford Cars, and Stanford Dogs datasets through comprehensive experiments.