Enhancing Tea Leaf Disease Classification with Cross- Attention Fusion and Magnitude-Aware Linear Attention

Abstract

Accurate tea leaf disease classification in real-world scenarios is hindered by complex backgrounds and the loss of fine-grained lesion details during CNN down sampling. To address this, we propose ResNet50-Dual-Fusion. It integrates a Cross-Attention Feature Fusion module (CAmodule) to adaptively reconstruct tiny lesion edges via cross-spatial interaction between shallow and deep features. Furthermore, a Magnitude-Aware Linear Attention (MALA) module with 2D Rotary Position Embedding (RoPE) is introduced to rectify magnitude neglect, effectively suppressing background noise. Evaluated on a 5,276-image dataset, our model achieves 85.96% accuracy (+3.00% over the baseline), outperforming architectures like ViT and Swin-Tiny. Grad-CAM visualizations confirm its superior lesion localization, providing a robust paradigm for automated crop disease diagnosis.