Research On Fast Defect Detection Method Of Printed Fabric Based On Machine Vision

With the development of economic conditions and standard of life,people's requirements for the appearance of shoes,clothing,home textiles and other textile products are increasing.Printed fabrics have many applications in the manufacture of modern textiles.However,in the manufacturing of printed fabrics,due to factors such as equipment failure,production environment,and improper human operations,it is easy to cause defects on the surface of printed cloth,which affects the economic benefits and industry reputation of textile enterprises.At present,most of the fabric defect detection in our country still depend on manual methods.Some companies have developed automatic defect detection systems,but most of them use traditional machine learning algorithms,which require manual design of defect features,which are not suitable for the detection of complex printed fabric surfaces.In recent years,deep learning has made great breakthroughs in industrial applications.This method can automatically learn features and has stronger feature representation capabilities.Therefore,this paper introduces deep learning models for defect detection,and designs the corresponding algorithm for the characteristics of printed fabric defects and the rapid detection requirements of the factory.This article first takes the survey in a textile factory workshop as a benchmark,and takes four types of defects with high frequency as the research object.Integrate defect characteristics,production environment,use cost and other factors,this paper carries out research on the selection of machine vision system components,choosing suitable light sources,cameras,lenses and other hardware,to build a system platform for printed fabric defects detection;and studies the principle of various types of deep learning algorithm to select target detection algorithm that is more in line with actual needs through the comparison of detection effects.Then,through the comparative study of the two ideas for object detection,the basic architecture on the basis of the single stage network YOLOv3 as the design prototype is selected,and an improved design is made based on the characteristics of the printed fabric defects.For the problem of the changeable shape of printed fabric defects,a deformable convolutional network is introduced to improve the effect of defect recognition;the sampling method of the original model is easy to ignore the problem of small defect information in the printed fabric,so the sampling structure is improved;for problem that factories need fast real-time detection,it is solved by the combination of shallow feature extraction network lightweight and deep network pruning;for the onestage model that has the problem of inaccurate detection of defect borders,the regression of the defect borders can be improved by additional generalized Io U loss calculation.Finally,an experiment is designed to verify the effectiveness of the printed fabric detection algorithm.In the preparation of the data required for training,the source data is collected and processed through the vision platform,camera SDK and developed image screenshot tool;the defect enhancement strategy is adopted to generate and enhance the data to solve the imbalance in the number of defects and insufficient diversity.In the experiment,the transfer learning strategy is used to reuse the knowledge of other fields to the prepared data domain to improve the convergence speed;the K-means algorithm based on Io U distance is used to cluster the real labels to obtain the suitable preset anchor for the data in this paper.After the experiment,results are obtained by the analysis of data,which demonstrate that the model in this paper can get an offline detection accuracy of 91.55%,the detection time is 0.026 s,and the detection performance is good;the online detection can run in real time at a speed of0.6m/s,and the detection accuracy reaches 86%,which can meet the accuracy and speed requirements of factory detection.