MSDF-Net: a cross-version lightweight detection framework based on deformable convolution and high-resolution feature...

Area of Science:

• Computer Vision
• Plant Pathology
• Remote Sensing

Background:

• Early detection of pine wilt disease (PWD) is crucial for effective forest management.
• Small, scattered early-stage lesions are difficult to detect amidst complex forest backgrounds and noise.
• Lightweight models often sacrifice pathological information due to feature simplification.

Purpose of the Study:

• To develop a lightweight object detection framework, MSDF-Net, for precise early identification of pine wilt disease.
• To enhance detection of small, sparse lesions by integrating advanced deep learning modules.
• To ensure the model's generalizability across different regions and pine species.

Main Methods:

• Proposed MSDF-Net, a lightweight object detection framework.
• Integrated a high-resolution P2 detection layer for small-target sensitivity.
• Utilized DCNv4 deformable convolution for adaptive spatial pattern modeling.
• Incorporated EMA attention mechanism for background suppression.
• Employed a dual-branch C2f DualConv module for multi-scale feature fusion.
• Evaluated the model on a cross-regional dataset.

Main Results:

• MSDF-Net achieved an mAP@0.5 of 80.1%, surpassing YOLOv8n by 5.1%.
• Demonstrated substantial improvement in early-stage disease detection (PWD-E) with a 20.2% AP gain.
• Maintained a low parameter count (2.67M) and moderate computational complexity (11.7 GFLOPs).
• Showed consistent performance improvements across YOLOv11n, YOLOv12n, and YOLOv13n versions.

Conclusions:

• MSDF-Net offers a generalizable and efficient solution for pine wilt disease detection.
• Its low parameter count and moderate complexity make it suitable for UAV edge deployment.
• The framework shows promise for real-time forest health monitoring applications.
• Further on-device validation is recommended for practical implementation.