Lightweight grading segmentation network for powdery mildew based on model pruning and knowledge distillation
Real-time, on-device segmentation of rubber tree powdery mildew is crucial for smart agriculture, yet existing deep learning models often fail to balance accuracy with computational efficiency. While large models can be precise, their high complexity hinders deployment on terminal devices; conversel...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772375525005507 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Real-time, on-device segmentation of rubber tree powdery mildew is crucial for smart agriculture, yet existing deep learning models often fail to balance accuracy with computational efficiency. While large models can be precise, their high complexity hinders deployment on terminal devices; conversely, smaller models often struggle to identify small, multi-scale disease spots in complex backgrounds. To bridge this gap, this study develops and validates light-PMSeg, a lightweight instance segmentation framework designed for accurate, real-time disease grading. Our contributions are fourfold: First, we establish a high-performance baseline, PMSeg, which integrates a Mixed Aggregation Network (MANet) and a Channel-wise Cross Fusion Transformer (CCFT) to significantly enhance sensitivity to small targets and multi-scale feature extraction. Second, to achieve on-device feasibility, we systematically compress PMSeg into light-PMSeg through a novel combination of slimming pruning and knowledge distillation (KD), which restores performance after parameter reduction. Third, we constructed a rubber tree powdery mildew segmentation dataset PM-Dataset containing 6,200 images. Finally, we propose an automatic grading algorithm based on the segmentation masks to provide actionable disease severity levels. Experimental results demonstrate the superiority of our framework. The final light-PMSeg model achieves a mAP50 (Mask) of 90.1% and an F1-score of 85.4% at a speed of 148.6 FPS. Crucially, this is achieved while reducing the model's parameters by 79.6% compared to the unpruned PMSeg, confirming its suitability for deployment on low-storage devices. This study delivers a validated, end-to-end solution that effectively reconciles high segmentation accuracy with a lightweight architecture, presenting a practical and robust method for intelligent powdery mildew monitoring. |
|---|---|
| ISSN: | 2772-3755 |