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

Optical remote sensing object detection faces challenges such as large variations in scale, slender and direction-sensitive targets, complex backgrounds, and limited deployment resources. This paper proposes a lightweight geometrically decoupled student network with a dynamic multi-teacher distillation framework. Based on YOLO11n, the student detector keeps the original classification branch while redesigning the regression branches at different scales. Lightweight regression towers are used for the shallow and deep branches, whereas a geometrically decoupled regression tower is introduced only at the intermediate branch to enhance localization for slender and orientation-sensitive objects with limited extra cost. A geometry-adaptive box loss is further employed to stabilize localization training. For knowledge transfer, three specialized teachers are constructed for semantic classification, geometric regression, and structural topology supervision. A branch-decoupled adaptive weighting strategy dynamically integrates their complementary knowledge for classification and regression distillation. Experiments on DIOR show that the proposed model reduces parameters by 6.8% and GFLOPs by 13.9%, while improving mAP50 by 0.23 percentage points over YOLO11n. Validation on NWPU VHR-10 and deployment tests using PT, ONNX, and TensorRT further demonstrate improved accuracy-efficiency trade-offs and practical inference acceleration.