| As raw materials in the field of medical and health,non-woven fabrics produce products closely related to our life.In the manufacturing process of non-woven fabric,due to the influence of production environment and manufacturing technology and other objective factors,the surface of non-woven fabric will produce black spots,white spots,fractures,scratches and other defects.At present,in most of the non-woven fabric production workshops in our country,surface defects are detected by manual,detection efficiency and accuracy are relatively low.For the past few years,deep learning network model has been successfully applied in the area of object detection,and will progressively displace the traditional machine vision to detect product defects.Therefore,in this dissertation,the surface defects of non-woven fabrics are detected based on the improved YOLOv3 algorithm,and the principal research contents are as follows:Firstly,a data set of surface defects of non-woven fabrics was constructed and the primitive YOLOv3 algorithm was analyzed.On the basis of the primitive YOLOv3 algorithm,surface defects of non-woven fabrics were detected.The experimental results showed that: The primitive YOLOv3 algorithm is not sensitive in the detection of small defect targets on the surface of non-woven fabric,there is a phenomenon of omission and the position of the generated prediction box is not accurate.Secondly,for the purpose of addressing the problem of missing detection and low recognition rate of little and medium defect targets in non-woven fabric surface defect detection,three improvements were made based on the primitive YOLOv3 algorithm combined with deep learning theory: Improved multi-scale feature prediction,and a new feature graph 104×104 was generated in the network model to extract more features of non-woven targets with small defects.The k-means ++clustering algorithm is used to promote the selection of cluster centers and produce the Anchor Box that is more consistent with the defect data set of non-woven fabric,and make the positioning of the prediction box more accurate.GIoU loss function optimization is carried out to substitute the positioning loss function in the primitive loss function with GIoU loss function,so as to better optimize the training model and obtain a network model with better convergence performance.On the basis of the previous work,three improved algorithms and the overall improved YOLOv3 algorithm are tested respectively,and the results show that: The overall m AP of the improved YOLOv3 algorithm is 88.05%,it is 7.96% higher than the primitive YOLOv3 algorithm.FPS is 25,which is also higher than the primitive YOLOv3 algorithm.The detection accuracy and detection speed are greatly improved. |
