Research on Fabric Defect Detection Technology Based on EDSR and Improved Faster RCNN

• Published in: Knowledge Science, Engineering and Management Vol. 13370; pp. 477 - 488
• Main Authors: Yao, Li, Zhang, Naigang, Gao, Ao, Wan, Yan
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: EDSR; Fabric defect detection; Faster RCNN
• DOI: 10.1007/978-3-031-10989-8_38
• ISBN: 3031109880; 9783031109881
• ISSN: 0302-9743

In order to solve the problems of various kinds of defects, defect ratio varies greatly, imbalanced defect aspect ratio and high integration degree with background in fabric defect detection, a method combining super-resolution reconstruction technology and deep learning detection was proposed. Firstly, the enhanced deep residual networks for single image super-resolution is used to enrich the defect feature information, reduce the fusion degree of defect and background texture, and enhance the extraction ability of various defect features. Then, the defect features are analyzed according to K-means clustering algorithm. Based on the three default anchor frame ratios provided by Faster RCNN, six new types of anchor ratios are added. Then, FPN module and DCNv2 module were introduced in Faster RCNN to improve the ability to identify defects with different areas and shapes. Finally, the pooling mode of ROI layer was modified to eliminate the error caused by quantization operation. The results of the three kinds of comparative experiments show that the method based on EDSR and improved Faster RCNN has a better overall recognition rate for multiple kinds of fabric defects than other current methods, and can be used in the production and operation of textile enterprises.