Detection and comparison of single and multiple defects in fabric patterns using machine learning methods

In the textile industry, quality control processes are very important for production efficiency and product quality. Defect detection with traditional methods is difficult, especially in nonwoven fabrics, due to small and irregular defects. An innovative deep learning-based solution to the problem of defect detection in nonwoven fabrics is presented by this study. The use of Artificial Intelligence-Based Object Detection (OD) and Transfer Learning (TL) methods aims to provide automatic and high-accuracy classification of fabric defects in this case. In the study, an original dataset labeled on the Roboflow platform and captured by us was used. In the dataset consisting of a total of 1239 images, four different defect classes (yag_damlasi, LYC, igne_kirigi and patlak) and 2584 labelings, an average of 2.1 defects were found. In the Object Detection approach, YOLOv5, YOLOv8, YOLOv11 and YOLOv9 models were used, while in the Transfer Learning approach, VGG16, ResNet50, EfficientNet_B0, DenseNet121, MobileNet_V2 and Inception_V3 models were preferred. The models were evaluated with various metrics including accuracy, precision, sensitivity and F1 scores. YOLOv5m demonstrated a notable performance with a confidence level of 0.244 and an F1 score of 0.81 across all classes. YOLOv8m achieved a confidence level of 0.268 and an F1 score of 0.75, YOLOv9-c a confidence level of 0.286 and an F1 score of 0.73, and YOLOv11- m a confidence level of 0.290 and an F1 score of 0.74. Transfer Learning (TL) models exhibited high performance on the curated dataset with a single defect per image, achieving test accuracies between 99% and 100%; VGG16, DenseNet121, and MobileNet_V2 stood out with 100% accuracy, while ResNet50 recorded the lowest result at 99.04%. However, on the original dataset with multiple defects per image, the performance of TL models declined, yielding accuracy rates of 87.30% for VGG16, 77.38% for EfficientNet_B0, 57.54% for Inception_V3, 47.22% for ResNet50, and 37.30% for MobileNet_V2. The experiments were conducted using a system where fabrics were passed through an inspection machine, enabling real-time analysis by the models. This study has shown that using deep learning techniques to identify defects in nonwoven fabrics is useful. It also aims to develop these techniques to assist in quality control procedures in the textile industry. In the future, testing the system on various types of fabrics and adding more defect classes can help make the method more generalizable. Keywords: Deep Learning, Nonwoven Fabrics, Defect Detection, Object Detection, Transfer Learning, YOLO Models