Aggregate attention module for fine-grained image classification

According to huge intra-class diversity and inter-class differences, fine-grained image classification has been a difficult topic for a long time. Attention mechanism has proven to be useful to aggregate features and discover discriminative local details. However, one issue is that the increasing parameters lead to unnecessary computation. In this paper, an attention mechanism named aggregate attention module is proposed to classify fine-grained images accurately with fewer parameters. Specifically, to balance the trade-off between performance and complexity, the proposed attention module combines channel attention with spatial attention in parallel, which effectively learns the key features and can be extended to other neural models easily. Meanwhile, we design cross-channel loss to explore discriminative fine-grained categories. Compared with state-of-the-art models, experiments reveal that our proposed model can achieve superior accuracy using different fine-grained image benchmarks (CUB-200-2011, FGVC Aircraft and Stanford Cars). To verify the effectiveness, we further evaluate our method through ablation study and visualization.