To alleviate the problems of label rewriting,unbalanced label assignment and feature coupling faced by YOLOv7 in detecting personal protective equipment,an improved YOLOv7 detection method is proposed.Firstly,the large-scale and medium-scale output layers of YOLOv7 are removed to reduce the label rewrite rate and to ensure that the output layer is adequately trained;secondly,the localization and classification of the output layer are decoupled to avoid that the feature representations of the different tasks affect each other and choose to detect the protective clothing at the bounding-box level,and the protective cap and protective gloves at the key-point level;Finally,a partial convolution is introduced to achieve real-time detection.In order to verify the effectiveness of the method,the proposed method is validated using image data of experimenters wearing protective equipment.The results show that compared with YOLOv7,the method improves the precision and recall by 4.1 and 4.5 percentage points,respectively,and the FPS is improved by 1.3 frames,which can satisfy the needs of personal protective equipment wearing detection in laboratory scenarios.
personal protective equipmentdetection of wearingYOLOv7single-scale outputfeature decouplingpartial convolution
NETL
NSTL
万方数据
