The Application Of Convolutional Neural Network In Textile Defect Detection

In recent years,deep learning technology has achieved great success in speech recognition,intelligent monitoring,face recognition,etc.Convolutional neural networks,as one of the main algorithms for deep learning,have also been widely used in industrial image defect detection.There are two main ideas for convolutional neural networks for fabric defect detection.One is to classify only the defects,and the other is to determine the defect category and locate the position of the defect.In this thesis,the three categories of gray fabric,yarn-dyed fabric,and complex stripe fabric are taken as our research target,the application value of the convolutional neural network in intelligent classification and detection of fabric defects is analyzed.Aiming at the difficulty of sample acquisition and the small sample size existing in the actual application scenario,the deep learning method is used to effectively classify and detect fabric defects.The research content mainly includes the following aspects:(1)This thesis combines transfer learning with a convolutional neural network model.Firstly,load the models AlexNet and GoogLeNet that have been pre-trained on the million datasets,and then fine-tune the parameters of the pre-trained model with the constructed data samples.The method greatly improves the training efficiency on the basis of shortening the training time and reducing the experimental hardware requirements.Because the convolutional neural network has different convolutional layers,which means that their receptive fields are different,the features extracted by AlexNet and GoogLeNet are merged on the basis of transfer learning.Finally,the SVM classifier is added for defect classification.The experimental results show that combination method of feature fusion and transfer learning is better than the application of AlexNet and GoogLeNet for separate transfer learning,and the recognition accuracy is increased by 4.8% and 10% respectively.(2)The defect detection method of yarn-dyed fabric based on YOLOv3.In order to classify and locate textile defects at the same time,the YOLOV3 network is used for color fabric defect detection.Different from the original YOLO method,the network uses Darknet-53 as the basic network.The improvement of the basic network is very eximious to reduce the pressure on the amount of calculation;the classifier uses multiple logistic classifiers;at the same time,it increases the multi-scale prediction.It solves the problems of low detection accuracy,low efficiency,and other problems in small target detection.The experimental results show that after testing on two types of yarn-dyed fabrics,the average accuracy of recognition is 98.07% and 97.18%.(3)Construction of a fabric defect detection platform.In order to further apply the ideas of this thesis to the actual industrial production,a fabric defect detection platform is constructed by using a fabric transmission mechanism,a light source,an image acquisition module,an industrial computer,and the like.Then combined with the detection algorithm in the thesis,a complete detection system is designed.At last,we look forward to the future development trend of the subject.There are 60 pictures,15 tables,and 74 references.