Research On Surface Defect Detection Of Polyester Fiber Fabrics Based On Deep Learning

With the continuous improvement of people’s quality of life,textiles made of polyester fibers and other materials have become indispensable items in people’s daily lives.As the demand for fabrics in the textile industry increases,the requirements for fabric quality have increased.In order to adapt to this requirement,the production and processing of textiles are developing in the direction of intelligence,and defect detection of textiles is a key step.At present,the manual detection method commonly used,its speed and accuracy are difficult to meet the needs of production,and the rapid development of computer vision technology provides a more intelligent and efficient technical means for textile defect detection.In view of the characteristics of unbalanced types of fabric defects,small proportion of defect area,imbalance of defect aspect ratio,and high degree of fusion between defect and background,it is difficult to apply to the complex characteristics of fabric defects by conventional artificial design feature expression,and this paper studies the fabric defect detection method based on deep learning.Improved feature extraction method based on multi-strategy.Aiming at the problems of small proportion of defect area and imbalance of defect aspect ratio,this paper adopts a multi-scale feature extraction method based on residual structure,realizes the fusion of image multi-scale features through FPN structure,introduces an attention mechanism in the backbone network,focuses on key channel features,and introduces deformable convolution operation in the backbone network to enhance the extraction ability of fabric defects in the network.Aiming at the problem of high degree of fusion between defect target and background,a two-stage pooled background suppression algorithm is added to the shallow layer of the backbone network of the detection framework,which can suppress the background to a certain extent,thereby strengthening the characteristics of the defect area.Defect detection method based on improved Faster RCNN.The improved defect feature extraction network is used to design a feature detection algorithm for fabrics.Firstly,the fabric defect data is clustered by K-means algorithm,and in order to adapt to the characteristics of unbalanced aspect ratio of fabric defects,a new parameter anchor is designed,and the proportion of 3 anchor boxes in the regional candidate network is expanded to 9,so as to improve the size of the anchor boxes generated in the regional candidate network and improve its ability to discriminate various defects.Faster RCNN’s backbone network,Res Net50,was then improved to accommodate the characteristics of the defect.Finally,in order to solve the quantitative operation error caused by the Faster RCNN network itself rounding the position of the feature map,the ROI pooling in the ROI module is replaced with ROI Align,and the position of the pixel is corrected,so as to reduce unnecessary errors.The last but critical step is to verify the proposed method based on three types of experiments.With a common defect dataset in 20 types of fabric defect identification being the experimental subjects,the first group is to verify the improvement of detection performance based on Faster RCNN.The second is to verify the change in the detection accuracy of various defects before and after the two-stage background suppression algorithm is adopted.And the third is to compare the algorithm proposed in this paper with several other commonly used algorithms.Based on the comparative experiment,the overall identification rate of 20 types of fabric defects is much higher,reaching 84.7% m AP,demonstrating that the method put forward in this paper is more appropriate for the production practice of textile enterprises.