Feature Re-Attention and Multi-Layer Feature Fusion for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) distinguishes sub-categories under a large category, with significant differences within classes and minor differences between classes. Many current methods propose to apply the attention mechanism to mine the most salient parts of objects to obtain more refined feature representations. However, using the attention mechanism will bring two limitations: One is that the attention mechanism usually pays attention to the most salient parts of the object and ignores the insignificant but discriminative parts. The other is that the separate use of features of specific object part ignores the connections between different parts. To address the first limitation, we propose a feature re-attention module (FRAM) to obtain feature representations of multiple enhanced specific object parts. At the same time, the multi-layer feature fusion module (MLFFM) is used to fuse the feature representations of multiple enhanced object-specific parts to learn semantically complementary information from each other. The proposed method can be trained end-to-end and does not require additional human annotation information. Extensive experiments on three fine-grained public datasets demonstrate that our method achieves state-of-the-art performance.