Textile Defect Detection And Classification Based On Deep Learning

The defects of textiles can directly affect the quality of products and the economic benefits of textile factories.The detection and classification of textile defects is important in the production process of the textile industry.At present,most factories still use skilled workers to check the quality of textiles.Using computer vision technology instead of manual can avoid the disadvantages such as slow detection speed,strong subjectivity of judgment standards,and high rate of detection missing.However,the current textile defect detection and classification methods belong to supervised learning,which requires a large number of defect images to train the network,but the appearance of defect images is low in the actual production process.In view of this situation,this article analyzes and researches the textile defect detection and classification,and proposes a textile defect detection method and classification method based on deep learning.This method only needs a small number of defect images to train the network.The main research content could be concluded by the following two aspects:1)Detection of textile defects.This dissertation proposes a defect detection model combining denoising autoencoder(DAE)and convolutional neural network(CNN).First,use normal images to train the denoising autoencoder,extract defect features,and then use CNN for defect detection.The DAE-CNN model uses a large number of easy-to-obtain normal images and a small number of defect images as training set.Experiments have proved that the accuracy and F1-Score of the DAE-CNN model are better than the commonly used VGG16 model,and the model proposed in this dissertation costs less calculations,which greatly improves the speed of defect detection.2)Classification of textile defects.CNN is a common classification model,but when the samples are not balanced,the accuracy of the CNN model in rare categories classification is low.Since Random Forest(RF)is not easily affected by sample imbalance,this dissertation proposes a defect classification model combining CNN and random forest.The CNN-RF model can improve the accuracy of rare categories while ensuring the accuracy of common categories.Experiments have proved that the accuracy and F1-Score of the CNN-RF model are better than the CNN model,and the CNN-RF model greatly improves the accuracy of rare category classification.