CBAM-Enhanced MobileNetV3-Small for Lightweight Multi-Class Crop Disease Recognition
Year: 2026
Type: conference-paper
Abstract: Lightweight crop disease recognition is important for deploying image-based diagnosis models on mobile terminals and edge agricultural devices. In this study, MobileNetV3-Small is used as the base network because of its low computational cost, and the Convolutional Block Attention Module (CBAM) is inserted into its bottleneck blocks to improve feature refinement. The motivation is that leaf disease symptoms usually occupy limited regions and may differ only in subtle color or texture patterns, which can be difficult for a compact backbone to represent sufficiently. By combining channel attention and spatial attention, the modified network strengthens lesion-related feature responses while largely retaining the efficiency of MobileNetV3-Small. Experiments were conducted on the PlantVillage dataset with 39 crop disease and healthy categories. The proposed MobileNetV3-Small + CBAM model achieved 98.46% accuracy, 98.29% macro-F1, 98.44% weighted-F1, and 0.9942 macro-AUC. Compared with the original MobileNetV3-Small, the accuracy increased by 2.31 percentage points, while the parameter size rose from 5.4M to 5.6 M and FLOPs increased from 60 M to 70 M. These results suggest that adding CBAM improves the recognition performance of MobileNetV3-Small in this PlantVillagebased setting with only a limited increase in model complexity.
Institution: Neusoft Institute Guangdong
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7201849784