Foreground Feature Enhancement and Peak & Background Suppression for Fine-Grained Visual Classification

Fine-grained visual classification task is challenging due to large variation in the appearance of the same subcategories and similarity in the appearance of different subcategories. To tackle the challenges, locating multiple discriminative regions plays a critical role. However, most previous works ignore the impact of background, which may provide negative clues that are not necessary or harmful for the network to classification. In this paper, we propose Foreground Feature Enhancement (FFE) module and Peak & Background Suppression (PBS) module, which are inserted in different layers of the CNN. The FFE module is designed to enhance and extract the most discriminative feature in the feature maps, and the PBS module is employed to suppress the peak features and background noise in the feature maps, forcing the network to mine other equally important features. Our proposed method can be trained end-to-end and does not require bounding boxes/part annotations. The experimental results achieve competitive performances on CUB200-2011, FGVC Aircraft, and Stanford Cars datasets.