Machine vision-based defect detection method for sewing stitch traces

Objective In order to solve the problems of slow speed, low efficiency, and high cost in conventional manual quality inspection methods for sewing thread, this study proposes a machine vision-based method for sewing thread defect detection in seams. This study aims to achieve fast, accurate, and automated identification of common defects such as cast thread, jumper thread, and broken thread in seams. This study also highlights the importance and necessity of improving product quality and production efficiency in the textile and garment industry.Method This study adopts a two-step approach for defect detection. Firstly, a low-cost array camera was adopted to capture real-time images of the sewing seam and the DeblurGAN-v2 method was employed to remove motion blurriness from the images, aiming at improving image clarity. Secondly, the student-teacher feature pyramid matching method was applied for anomaly detection, which transfers the knowledge from a pre-trained ResNet-34 model as the teacher network to a student network with the same architecture, so as to learn the distribution of normal images. By comparing the differences between the feature pyramids generated by the two networks as a scoring method, the defect detection system made decisions on whether the image has anomalies, and marked the abnormal areas with a heat distribution map.Results The defects of flat stiteh fabric and overstitch fabric were tested and the performance of the proposed method was evaluated in terms of recall and accuracy rates. The results show that the proposed method can effectively detect various sewing thread defects and has high recall and accuracy rates for different types of defects. This study also provided some examples of defect detection results and scores for different types of defects. Conclusion The feature pyramid matching technique is applied in the field of stitch trace detection. By adding the difficult sample mining technology, the average detection accuracy is increased to more than 95%, and the detection speed of a single image is less than 0. 04 s. Aiming at image motion blur ring caused by jitter and fast movement. The DeblurGAN-v2 framework is used as the framework of deblurring algorithm, and the blueprint convolution is added to change the backbone network, and the processing speed of a single image is kept below 0. 06 s. The model has excellent interference resistance and high processing speed, and can meet the requirement of stitch trace recognition. © 2024 China Textile Engineering Society. All rights reserved.