Feature Channel Adaptive Enhancement for Fine-Grained Visual Classification

Fine-grained classification poses greater challenges compared to basic-level image classification due to the visually similar sub-species. To distinguish between confusing species, we introduce a novel framework based on feature channel adaptive enhancement and attention erasure. On one hand, a lightweight module employing both channel attention and spatial attention is designed, adaptively enhancing the feature expression of important areas and obtaining more discriminative feature vectors. On the other hand, we incorporate attention erasure methods that compel the network to concentrate on less prominent areas, thereby enhancing the network’s robustness. Our method can be seamlessly integrated into various backbone networks. Finally, an evaluation of our approach is conducted across diverse public datasets, accompanied by a comprehensive comparative analysis against state-of-the-art methodologies. The experimental findings substantiate the efficacy and viability of our method in real-world scenarios, exemplifying noteworthy breakthroughs in intricate fine-grained classification endeavors.