Research On Substation Construction Safety Belt Segmentation Method Based On Double Attention Mechanism Of Space And Channel

"Electrical Safety Work Regulations" stipulates that construction personnel must use safety belts when constructing above 2m from the ground.Wearing safety belts can effectively protect the safety of construction workers.This article takes the construction personnel,safety helmets and safety belts in the image of the substation construction site as the research objects.Using deep learning methods,the following tasks have been completed:Using deep learning to perform semantic segmentation of substation construction site images,first a substation construction site image semantic segmentation data set that can be used for deep model training is needed,but there is currently no public data set suitable for this problem.Therefore,the labeling tool Labelme was used to label the substation construction site pictures,and a sample substation construction site image semantic segmentation data set was constructed through sample amplification.The dataset includes four categories: construction workers,safety belts,hard hats,and backgrounds.The deeplab V3 network with high segmentation accuracy is selected to perform semantic segmentation on the image of the substation construction site,and the network is trained by transfer learning.After the segmentation is completed,the method of detecting whether the substation constructors are wearing safety belts and hard hats through the segmented images is introduced.Experiments show that the deeplab V3 network can perform end-to-end semantic segmentation on the image of the substation construction site.The average pixel classification accuracy can reach 90.9%,and the average intersection ratio can reach 67.8%.In view of the deeplab V3 network’s segmentation accuracy of the construction personnel,safety helmets and safety belts in the image of the construction site of the substation,the segmentation of the construction personnel and safety belts,especially the safety belts,will have missing edges.Channel and spatial dual attention mechanisms are added to the residual unit and the Atrous Spatial Pyramid Pooling(ASPP)module of deeplab V3 network to optimize the segmentation effect.Experiments show that the deeplab V3 network with the dual attention mechanism of space and channel is 1.6%higher than the previous average pixel classification accuracy,and the average intersection ratio is 4.1% higher.The segmentation details have become richer and the seat belt segmentation has become smoother.The boundary pixel classification is more accurate.