Deep learning for fabric defect detection

• Main Author: Nangia, Saniya
• Other Authors: Zheng Jianmin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Deep learning; Ensemble models; Computer vision; Defect detection
• Online Access: https://hdl.handle.net/10356/175318
• Institution: Nanyang Technological University

Defects occur in the fabric production process due to machine-related errors or system faults. Deep learning solutions can be employed to increase the frequency and accuracy of detection, while reducing the time, cost and labour involved in textile manufacturing. While traditional models like convolutional neural networks can identify larger defects on solid-coloured fabrics, they are unable to accurately identify smaller defects, especially on a variety of patterned and textured fabrics. Single-stage models and transformer-based object detection models have resulted in vast improvements in model accuracy and efficiency. In this study, a modified StackBox ensemble method is implemented. The base models, YOLOv8 and RT-DETR, are fine-tuned on a dataset of patterned fabric defects to find the optimal model configurations. Their confidence scores are adjusted to reflect single or dual votes for predicted bounding boxes. Finally, a Support Vector Machine Regression meta-model is chosen to combine the predictions of both models, attaining a higher average precision and average recall than baseline models and other ensemble methods. A comparison is provided with Non-Maximum Suppression (NMS), Soft-NMS and Weighted Boxes Fusion ensemble techniques. The performance of models such as Cascade R-CNN, Faster R-CNN and RetinaNet is also analysed in comparison to the base models and modified StackBox method.