A Tobacco Foreign Object Detection Method Based on Intermediate Feature Layer Distillation

Foreign object detection in tobacco production is critical. To address the challenge of balancing high detection accuracy and real-time performance, this study proposes an object detection framework based on intermediate feature layer distillation to enhance model performance. Through attention-guided distillation and non-local distillation, the intermediate layer knowledge of the teacher model is transferred to the student model, thereby improving its accuracy in detecting non-tobacco foreign objects in tobacco production. Experimental results on the tobacco foreign object dataset show that, compared to direct training, the proposed method improves the mAP of the PAA50 model by 2.5% and the AR by 1%. The feature-based knowledge distillation method proposed in this study effectively improves the performance of tobacco foreign object detection models, thereby enabling the efficient detection of non-tobacco foreign objects in tobacco production.