Lightweight Safety Helmet Wearing Detection Method Based on KD-YOLO
The detection of safety helmet wearing is crucial for safety management on construction sites.Exist-ing detection models suffer from slow detection speed,low detection accuracy,and difficulties in on-site de-ployment.Accordings,a lightweight safety helmet detection model,KD-YOLO,based on YOLO5s is pro-posed.This model incorporates the concept of knowledge distillation,using an improved lightweight network as the student network,to enhance the student network's detection accuracy through teacher-student knowl-edge transfer.It employs the lightweight attention module,Shuffle Attention,to enhance the focus on useful information of the images.Additionally,it employs the Wise-IoU bounding box loss function based on a dy-namic non-monotonic focusing mechanism to balance anchors of different qualities,reducing false positives and false negatives.Experimental results show that the improved lightweight safety helmet detection model,optimized through knowledge distillation and other strategies,achieves a 3.5%increase in mAP,a 18.1%re-duction in parameter count,and a 1.8%reduction in floating-point operations,respectively.This model meets the real-time safety helmet wearing detection requirements on construction sites,while consuming less memo-ry and being easy to deploy on the construction sites.