Research On Fabric Defect Detection Method Based On Machine Learning

With the progress of science and technology,artificial intelligence has been widely applied in the manufacturing industry,in order to enhance the core competitiveness of our equipment intelligence.Machine learning is an important branch of artificial intelligence.In industrial automation production,intelligent detection technology based on machine learning has been widely used in many industries.Among them,the textile industry is the pillar industry of national economy and social development in our country,effective detection and control of textile surface quality is one of the key links for modern textile enterprises to control cost and improve product competitiveness.Automatic fabric defect detection algorithm based on machine learning has become a hot topic in the cross research field of textile equipment and information discipline.Due to the variety of fabric defects,different shapes,complex texture and rich color of fabric,there are many problems in the detection process of fabric defect detection algorithm.Therefore,solving these common problem of few fabric defect samples and insufficient labeled samples,uneven defect sample classification and small pixel defects,and how to balance accuracy and complexity of deep learning network model in surface defect detection of raw fabric,patterned fabric and yarn-dyed fabric have become the key to research fabric defect detection methods based on machine learning.Aiming at these problems,this paper has carried out in-depth research,including the following aspects;(1)Aiming at the problems of low detection rate and long running time of traditional fabric defect detection methods,a fabric defect detection method based on positive sample integral image Elo-rating was proposed.In this method,the defect-free image is divided into blocks based on the competition mechanism in sports events,and different blocks are regarded as multiple players participating in the competition.The corresponding grade is updated by the result of the competition among the contestants,and the defect region and texture region are distinguished by the high and low status of the Elo grade,and finally the defect detection result is obtained.At the same time,in order to improve the detection efficiency of the system,the idea of integral image is added on the basis of the Elo-rating algorithm,which saves the detection time.The results show that the method has a higher detection rate and better detection effect for the defects of raw fabric,yarn-dyed fabric and patterned fabric,and the detection time is significantly shortened.Especially for star-patterned fabric,the average detection time per image is about 24.18 seconds less than that of the original Elo-rating algorithm.(2)Aiming at the problems of few fabric defect samples and insufficient labeled samples,a fabric defect detection method based on unsupervised sparse dictionary learning was proposed.In this method,the normal defect-free fabric image is taken as the learning object,and the feature column is randomly selected from the defect-free image as the random dictionary.In this way,the texture,edge and background information of the image can be displayed to a greater extent,and the shallow detail features of the image can be more balanced.The final learning dictionary is obtained by Orthogonal Matching Pursuit and K-Singular Value Decomposition algorithm.Because the learning dictionary cannot approximate the defect area well,the reconstructed error matrix can achieve the purpose of enhancing the outliers and effectively improve the effect of defect detection.The performance of the proposed algorithm is evaluated on datasets of raw fabric,yarn-dyed fabric,and patterned fabric.Experiment results show that the proposed method achieves good detection results on three datasets.(3)Aiming at the decreased detection rate problems when uneven fabric defect sample categories and small pixel defects,a deep supervised UNet++fabric defect detection method with pixel level pruning is proposed.In this method,the features of the convolutional neural network are connected by cascading operation,and the shallow detailed features and deep features are fused.The idea of deep supervision is added to make statistics of the output of different depth models,so as to conduct reasonable pruning according to the texture,color,defect and other characteristics of different kinds of fabric.Deep supervision provides a better trade-off between network depth,speed,and accuracy.Finally,the Median Frequency Balanced weighted Cross Entropy loss function is introduced to solve the decreased detection rate problems when uneven fabric defect sample categories and small pixel defects.The experimental results show that the average detection rate of this method is 97.68%and 99.01%for raw fabric and patterned fabric,respectively.(4)Aiming at the problems of how to balance the accuracy and complexity of deep learning model and how to reduce the deployment cost of model,a lightweight network fabric defect detection method for edge equipment was proposed.This method integrates a convolution block attention module into the Backbone part of the lightweight network YOLOv5s,which makes up for the deficiency of the network’s ability to extract information and reduces the loss of contextual information in high-level feature maps.A feature enhancement module is integrated into the Neck part to enhance the representation of the feature pyramid and improve the reasoning speed of model.At the same time,the loss function of YOLOv5s is modified to CIoU Loss by considering the overlapping area,center point distance and anchor frame aspect ratio,which improves the speed and accuracy of the prediction frame regression.The raw fabric and patterned fabric datasets are tested on the deep learning workstation platform.Compared with other two-stage and one-stage detection methods,the proposed method can make a good trade-off between model accuracy and complexity,and meet the real-time requirements of industrial sites.After that,the trained model is transplanted to edge device NVIDIA JetsonTX2,and the model is optimized by TensorRT to test the real-time performance of the model deployed in the industrial field.