A Lightweight Student Network with Dynamic Multi-Teacher Distillation for Optical Remote Sensing Object Detection

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Optical remote sensing object detection faces challenges including scale variation, slender targets, and complex backgrounds.
• Limited deployment resources necessitate efficient and accurate detection models.

Purpose of the Study:

• To propose a lightweight, geometrically decoupled student network with dynamic multi-teacher distillation for improved optical remote sensing object detection.
• To enhance the localization accuracy of slender and orientation-sensitive objects with minimal computational overhead.

Main Methods:

• A lightweight student network based on YOLO11n, featuring redesigned regression branches and a geometrically decoupled intermediate branch.
• A dynamic multi-teacher distillation framework with specialized teachers for classification, regression, and topology, using a branch-decoupled adaptive weighting strategy.
• Implementation of a geometry-adaptive box loss for stable localization training.

Main Results:

• The proposed model achieved a 6.8% reduction in parameters and a 13.9% decrease in GFLOPs compared to YOLO11n.
• An improvement of 0.23 percentage points in mAP50 was observed on the DIOR dataset.
• Validation on NWPU VHR-10 and deployment tests demonstrated enhanced accuracy-efficiency trade-offs and practical inference acceleration.

Conclusions:

• The developed lightweight network effectively addresses challenges in optical remote sensing object detection.
• The dynamic multi-teacher distillation framework significantly improves localization accuracy and model efficiency.
• The model offers practical advantages for real-world deployment due to accelerated inference and reduced resource requirements.