ENet-CAEM：改良型EfficientNetB0とマルチスケールアテンション機構に基づくイチゴ病害同定モデル

Real-time diagnosis of strawberry diseases plays a key role in sustaining yield and improving field management. However, achieving reliable recognition remains challenging. Lesions often display irregular shapes and appear at different scales, which complicates detection. Field images also contain cluttered backgrounds, while many diseases look visually alike, making differentiation more difficult. In addition, collecting data under real conditions is not easy, resulting in small datasets on which deep learning models tend to overfit and fail to generalize.

To address these issues, this study introduces ENet-CAEM, a redesigned EfficientNetB0 framework equipped with modules tailored for disease recognition. The Channel Context Module helps the network capture key lesion features while suppressing background noise. The Multi-Scale Efficient Channel Attention module applies multiple one-dimensional filters of varying sizes in parallel, enabling the model to highlight critical patterns, tell apart similar diseases, and adapt to lesions of different scales. A lightweight version of Atrous Spatial Pyramid Pooling is further integrated, allowing the network to perceive features at multiple spatial ranges. To balance local detail with global context, a mixed pooling strategy is adopted, enhancing robustness when lesion shapes change. Finally, Learnable DropPath and label smoothing are applied as regularization strategies, reducing overfitting and improving generalization on limited data.

Experiments show that ENet-CAEM achieves 85.84% accuracy on a self-built dataset, outperforming the baseline by 4.29%. On a public strawberry dataset, the model reaches 97.39%, surpassing existing approaches.

The proposed ENet-CAEM model shows superior accuracy and robustness over existing methods, providing an effective solution for strawberry disease recognition in practical field environments.