Research On Fabric Surface Defect Detection Based On Transfer Learning

With the rapid development of the textile industry and the improvement of people's life,the quality evaluation of the fabric has gradually become the focus of attention in the textile industry.The current detection algorithms have poor universality,which means that one detection algorithm can only be applied to one kind of fabric.It causes the more difficulty and cost in the development of fabric detection software.Aiming at this situation,the transfer learning method is applied to fabric defect detection,which not only overcomes the poor universality of current fabric defect detection algorithms,but also is suitable for the detection of small sample datasets.The main work contains the following aspects:(1)The research background and significance of fabric surface defect detection are introduced,and the research on fabric surface defect detection at home and abroad is divided into three stages: based on traditional image processing algorithm,traditional machine learning algorithm and deep learning on the large dataset algorithm.Aiming at the shortcomings in the current deep learning algorithm,the fabric defect detection algorithm based on transfer learning is proposed.(2)Combining the transfer learning with deep learning,the defect detection and classification of electronic grade glass fiber cloth based on deep transfer learning is used.This method mainly uses the idea of transfer learning to improve the deep learning model ResNet,whose feature extraction part is retained,having the same feature extraction ability as the original model.A new classifier consisting of a normalization layer,a DropOut layer,an activation layer,a fully connected layer,etc.,is designed.Combining the feature extraction part and the new classifier,a new network is obtained.By applying the network,the electronic cloth defect dataset is trained and tested,and the detection accuracy is 99.1%.In addition,the real-time requirements of the industrial site are also met.(3)Aiming at the problem of poor universality of the current fabric defect detection algorithms,a fabric defect detection algorithm based on improved DANN network is proposed,which combines transfer learning with deep learning and adversary learning.This method,based on DANN,enhances the feature extraction ability and solves the problem of poor performance of under the condition that the source domain and the target domain are not similar.The MMD-DANN network is designed by using the deeper ResNet50 network structure and the MK-MMD layer which is added to the original network to give different weights to the target domain features.Finally,the DANN network and the MMD-DANN network are tested on the white fabric defect dataset,the color fabric defect dataset,and the electronic cloth defect dataset.The experimental results show that compared with the DANN network,the improved MMD-DANN network has higher detection accuracy and efficiency on the three defect sample sets.There are 22 figures,7 tables and 82 references in this thesis.