Lightweight Hybrid Attention Network for Edge Deployment in Crop Disease Recognition

Early and accurate identification of crop diseases holds significant economic value for agricultural production. Although deep learning-based methods demonstrate potential, existing lightweight models face challenges in balancing parameter efficiency and multi-scale feature modeling capability under resource-constrained edge computing scenarios. To address this, this paper proposes a Lightweight Hybrid Attention Network (LHAN) for crop disease recognition, tailored for edge deployment. By replacing the Squeeze-and-Excitation (SE) module in EfficientNet-B0 with a parallel lightweight hybrid attention mechanism (LHA), the proposed model reduces parameter volume while enhancing multi-dimensional feature representation capability. Experimental results on the tobacco leaf disease dataset demonstrate that the EfficientNet-B0 integrated with LHA achieves a 12% reduction in parameter count compared to the baseline model while improving classification accuracy by 0.4%. This study provides a lightweight, efficient, and deployable solution for real-time disease diagnosis in agricultural edge computing scenarios.