An Efficient Attention-Enhanced MobileNetV2 Framework for Plant Disease Detection on Resource-Constrained Devices
TL;DR
The evidence supports a compact accuracy–cost compromise for the tested conditions, while field robustness, energy use, repeated training runs, and target-device behavior remain open validation requirements.
Abstract
Leaf disease diagnosis needs models that are accurate enough for agronomic use yet small enough for constrained computing settings. This study examines a late-attention MobileNetV2 design in which one Convolutional Block Attention Module (CBAM) is inserted between the last MobileNetV2 convolutional map and global average pooling. The experiments use 54,306 controlled-background PlantVillage images spanning 38 classes. Under a uniform saved-model re-evaluation, MobileNetV2 + CBAM obtained 97.17% accuracy and 97.15% weighted F1-score, whereas MobileNetV2 obtained 96.78% and 96.73%. On the converted models, paired testing gave a 0.64-percentage-point accuracy advantage for the CBAM variant (95% CI: 0.31–0.96; exact McNemar p<0.001). The proposed network has 4.02 million parameters, costs 0.604 GFLOPs (about 0.302 GMACs), and yields a 4.07 MiB dynamic-range-quantized TensorFlow Lite file with 96.70% accuracy. Batch-one inference on an Intel i7-11800H CPU with TensorFlow Lite/XNNPACK and eight threads reached a median of 45.42 ms (P95: 102.54 ms), excluding preprocessing. Grad-CAM inspection illustrates both lesion-centered activation and unresolved shared errors. The evidence therefore supports a compact accuracy–cost compromise for the tested conditions, while field robustness, energy use, repeated training runs, and target-device behavior remain open validation requirements.