Research On Defect Detection Method Of Non-woven Fabric Based On YOLOv5

In the medical and health field,non-woven fabrics have always been an important raw material for medical products and medical consumables,and are widely used in the production of surgical masks and protective clothing.Because it is impossible to ensure the absolute dust-free and aseptic working environment in the production,transportation and processing of non-woven fabrics,and the friction between nonwoven fabrics is easy to generate static electricity,static electricity itself is easy to absorb dust,and these impurities mixed on the surface of non-woven fabrics will pollute the surface of the mask.The defect detection of non-woven fabrics mainly depends on manual detection,which is inefficient.Therefore,it is necessary to develop a fast,accurate and efficient non-woven fabric defect detection system.In this paper,under the background of the above-mentioned topics,a defect detection method of non-woven fabrics based on YOLOv5 is studied.The main research contents of this paper are as follows:This paper analyzes the defect characteristics of non-woven fabrics in detail,puts forward the difficulties of non-woven fabric defect detection,systematically expounds the basic composition principle of convolutional neural network,summarizes the advantages and disadvantages of two kinds of algorithms in the field of target detection,and chooses YOLOv5 as the basic algorithm of non-woven fabric defect detection according to the particularity and superiority of the algorithm,and analyzes its existing problems and gives solutions.Aiming at the problem that the features extracted by convolution kernel are local and the ability to express global features is not strong,this paper puts forward YOLOv5-tf model,and replaces the backbone network of the original YOLOv5 network with Transformer.By fully integrating global features with a new architecture,the ability to identify small defects in non-woven fabrics is improved.In view of the fact that the original YOLOv5 network only simply superimposed the features at different scales,and did not fully consider the differences of features at different levels,this paper proposed YOLO V5-MSAM(Multiscale Attention Mechanisms)model,and designed a multiscale attention mechanism by using local feature information module and global information module,and applied it to the feature pyramid.The process of non-woven fabric defect detection is determined,and the experimental results are analyzed.The experimental results show that the m AP50 of YOLOv5-tf model is improved from 82.9% to 89.4% and 6.5%,and YOLOv5-tf model is superior to YOLOv5 in overall convergence speed and training stability.The value of m AP50 of YOLOv5-MSAM model is 9.4% higher than that of the original YOLOv5 model.YOLOv5-tf-MSAM model with two improved points is superior to the above three models.NCNN is used to deploy the non-woven fabric defect detection model on mobile devices,and the detection speed can be 40 FPS in GPU mode.