Research On Safety Helmet Wearing Detection Methods Of YOLOv4 Lightweight

In recent years,most of the safety accidents that occur in various construction sites are caused by workers not wearing safety helmets as required.Intelligent helmet wearing detection is very necessary for the safety of workers.The methods based on deep learning play an important role in safety helmet wearing detection,however,the object detection network is difficult to balance network depth,model precision,and model size,so that it is not easily applicable to real scenes.Therefore,this thesis investigates the task of safety helmet wearing detection based on YOLOv4 lightweight network and proposes the following solutions.1.This thesis proposes a lightweight network model DSB＿YOLOv4 for safety helmet wearing detection,which is obtained by improving the YOLOv4 backbone network with the DSBlock module.The DSBlock module replaces the standard convolution with depthwise separable convolution,and combines the compressed excitation mechanism module to increase the weight of feature information to ensure accuracy while lightweight.The backbone network of DSB＿YOLOv4 is DSBNet.It has five combinations and includes two modes: "2+2 mode" that two standard convolutions add one depthwise separable convolution."3+2 mode" that three standard convolutions add one depthwise separable convolution.In this network,each module is not stacked internally,which reduces the depth of the network,the number of parameters and model size.The validity of DSB＿YOLOv4 network is verified in the public datasets.2.This thesis further explore the DSBlock module.the standard convolution of the DSBlock was replaced by the depthwise separable convolution.The resulting module is called DSBlock2,and the corresponding DSBNet2 has five combinations,and also contains two modes: "2×2 mode" that is two depthwise separable convolutions,and "2×2+1 mode" that two depthwise separable convolutions add one dilated convolution.Similarly,in this thesis,the experiments verify the effectiveness of DSBNet2 corresponding to the DSB＿YOLOv4 network on public datasets.3.This thesis proposes a network,named RA＿YOLOv4,it is improved the feature aggregation network of YOLOv4 by reverse attention mechanism module for helmet wearing detection.The network reduced depth without loss of accuracy through increasing the saliency features of the object.The network takes the output feature information of the backbone network as the target item,and removes the object information from this target item,then aggregates it with other lateral outputs feature information of the backbone network.In this way,the network is allowed to discover the missing edge and detail information about the object,which leads the network to focus on the construction workers wearing helmets.The experiments on the current public datasets validate the effectiveness of the RA＿YOLOv4 network model without adding additional consume,and lightweighting can also be achieved by conjunction with an improved feature extraction backbone network.