Channel Boosting, Cross-Layer Feature Integration, and Multi-Scale Classification for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) is extremely challenging compared to basic-level classification. Although some attention-based methods have achieved promising results, they commonly fail to explore the benefits of multi-scale visual information and neglect the importance of information from intermediate layers in the classification network. In this work, we propose a unified framework for FGVC that integrates channel attention, spatial attention, and cross-layer feature interaction. We also demonstrate the significance of information from intermediate layers. Firstly, we introduce a channel boosting module that enhances the most salient parts in the feature maps. Secondly, we enhance and classify the outputs of intermediate layers, emphasizing their importance in the classification task. Furthermore, we introduce a cross-layer feature interaction module capable of integrating information from different feature layers. Lastly, the suggested framework does not require any bounding boxes or part annotations and can be trained seamlessly end-to-end. Extensive experiments show that our proposed method achieves competitive performance on three fine-grained classification benchmarks.