The Research Of Helmet Detection Based On Convolutional Neural Network

With the development of modern industrial production towards automation and intelligence,automatic identification technology has became an important direction of industrial production application,in which the automatic identification and detection of safety helmet has a special significance for production safety,and has gradually became an important application of deep learning intelligent detection direction.The traditional identification of helmet wearing is through monitoring for manual supervision,but the manual supervision time is long,the efficiency is low,and there are subjective differences,it is difficult to achieve real-time supervision and detection,can not meet the needs of production safety in modern society.With the support of deep learning technology,many scholars at home and abroad are committed to solving the problem of target detection through continuous experiments,and put forward a large number of target detection algorithms.Both in terms of real-time performance and accuracy,YOLOv3 algorithm performs very well.The algorithm uses the neural network of Darknet-53,with high classification accuracy,less network layers and fast calculation speed.It finds the prior frame by clustering the objects in the image,and completes the positioning of the object frame by regression,so that the target detection task can be completed quickly.This paper mainly introduces the research on the safety helmet wearing detection of the industrial production scene image by using the YOLOv3 algorithm and the improved YOLOv3 algorithm.The research work is as follows:1.Built the HELMET data set.The cameras are deployed on the construction site,and the video is processed into images.The similar images are deleted by removing duplicate images method,and the data set of safety helmet is constructed.The data set is expanded by adding noise,rotating and flipping the images,and the corresponding XML file is obtained after annotation.Finally,3174 images are sorted out,and they are divided into training set and test set by the ratio is 7:3.2.Trained and tested of YOLOv3 network.After getting the prior frame of the training set by clustering algorithm,the model of YOLOv3 algorithm is trained on the training set.After getting the model with the least loss,the model is tested on the test set,and the mean average precision of YOLOv3 algorithm on the test data set is obtained.3.An improved YOLO-CDF network is proposed and compared.It is observed that some pictures tested by YOLOv3 algorithm are difficult to get effective detection because of the change of human body posture,the small size of safety helmet in the picture and some ambiguous places in the picture.Therefore,an improved algorithm with dual attention mechanism and deformable convolution is proposed to improve the recognition and detection ability of the algorithm,which can effectively deal with the change of human body.After the ablation experiment and comparison with the classical target detection algorithm,the feasibility of the improved algorithm is proved by the detection results on the test set.4.An improved YOLO-DPN network is proposed and compared.It is noticed that gradients disappear when the layers of deep learning network deepen gradually.Res Net and Dense Net solve this problem by learning residuals and dense connections respectively.Based on this,we apply the network DPN which combines the two innovations to target detection,and make corresponding improvements to the YOLOv3 network.Through the comparative experiments,it is proved that the improved network model can make better use of the feature information and achieve better detection effect.