| In the textile industry,the presence of defects on the surface of fabric is an essential factor in determining fabric quality.Therefore,identifying fabric defects forms a crucial part of the fabric production process.The method based on deep learning has many advantages in the field of fabric defect detection,but there are still many limitation: 1)The defect feature of fabric has the characteristics of complex texture,multiscale,and extreme aspect ratio and so on,this leads to the problems of low recognition rate and high probability of missing and wrong detection in the current fabric defect detection method based on deep learning.2)The detection network trained on a small amount of labeled data has a weak ability to identify and detect defects of fabric,which leads to the low quality of pseudo-tags generated by the traditional semisupervised detection model,and poor classification and localization ability of defects of fabric.In view of this,this paper studies the semi-supervised fabric defect detection method based on convolutional neural network.The principal researches are as follows:(1)Multi-layer Feature Extraction with Deformable Convolution for Fabric Defect Detection: A multi-layer feature extraction algorithm is proposed to integrate the semantic information of high level features with the location information of low level features,so as to improve the detection accuracy of multi-scale fabric defects.The deformable convolution algorithm is used to enhance the generalization ability of the algorithm to deal with complex shape defects,and improve the feature extraction ability of defects with complex shape and extreme aspect ratio.Integrating the Roi Align and Cascade RCNN algorithms,by increasing the IOU threshold in stages,the problem of false positive results is reduced,and the accuracy of the fabric defect detection algorithm is increased.(2)Semi-supervised fabric defect method based on pseudo label: To solve the problem of low quality false labels generated by traditional multi-stage semi-supervised detection methods,an end-to-end semi-supervised detection method combined with EMA(Exponential Moving Average)strategy was adopted,both the teacher and student models were combined for learning.EMA strategy is used for the student model to update the teacher model,and fake labels can be mutually strengthened with the fabric detection mode.The image classification strategy based on reliability index is adopted,which takes credibility as a measure of whether pseudo label is the classification loss of defects,and solves the problem that the recall rate of fabric defect detection model decreases too fast due to the high prospect score threshold of traditional semisupervised detection methods.Since there is no strong positive correlation between the localization accuracy and the foreground score of the box candidates,the jitter strategy of the box candidates frame is proposed in this paper,and the regression variance of the box candidates is used to represent the reliability of the positioning accuracy,which improves the positioning accuracy of the semi-supervised detection method for fabric defects. |
