Safety Helmet Wearing Detection Based On Deep Learning In Complex Operation Scenarios

With the continuous popularization of the importance of safety production and the continuous development of deep learning field,the application of computer vision to production safety has become the mainstream trend.In the traditional industry,manufacturing and construction fields,the safety helmet wearing detection based on computer vision has gradually become a popular research direction.Due to the extremely complex operation scene of the safety helmet detection and dense personnel,the feature capture ability of the model is relatively high.At the same time,there are strong light irradiation or dark conditions,and the safety helmet target is small,so there is still a bottleneck in the safety helmet detection.Therefore,the more suitable for the complex operation of the construction site has important theoretical value and research significance.In this thesis,the YOLOv4 and YOLOv5 models are optimized and improved respectively,and the detection ability of the two models is improved in the complex work scene.The main work of this thesis is as follows:In order to solve the problems of low precision and strong interference of helmet detection in complex job scenes,a two-dimensional deep feature fusion helmet detection model is proposed in this thesis.The YOLOv4 model was improved as a baseline model,and a deep multi-scale feature fusion architecture was designed to capture the tiny helmet targets,and a Convolutional Block Attention Module(CBAM)was introduced into the model,a two-dimensional feature fusion channel is constructed to enhance the detection ability of the model for different types of targets,and the convolution layer of the head of the model is replaced by the depth separable convolution,thus realizing the efficient detection of the helmet.In this thesis,DCMS-YOLOv5 model is proposed for helmet detection,and the single-channel feature extraction method of YOLOV5 model is improved,and a dual-channel laterally connected backbone network is constructed,the feature image is entered into two channels in order for feature extraction and fusion,so that the model can better reduce the background noise interference in the detected image.The Neck structure is extended vertically,and the interaction channel between deep and shallow information is constructed by multi-scale vertical expansion,which enhances the ability of the model to detect the long distance.At the same time,add Focus slice to the model to improve the running speed of the model.This thesis conducted a validation analysis on the data set Safety Helmet Wearing-Dataset(SHWD),and compared with other helmet wearing detection model,The experimental results show that the deep feature fusion helmet detection model based on two dimensions and the helmet detection model based on DCMS-YOLOv5 can effectively improve the precision of helmet detection,and the effect of object detection in complex job scenes is obvious,has the strong generalization ability.