M. Gao, E. Shim, M. Zhu
North Carolina State University,
United States
Keywords: Defect Detection, Deep Learning, YOLOv8, Frequency-domain Features, Industrial Inspection
Summary:
Detecting surface defects in textile materials remains challenging due to the presence of small anomalies and complex textured backgrounds. This work focuses on deep learning-based defect detection methods tailored to the structural characteristics of woven materials. From a data perspective, the AITEX fabric defect dataset is reorganized through defect category redefinition, patch-based image cropping, and data augmentation to better support defect detection tasks. In parallel, a proprietary textile defect dataset based on real production samples is being constructed to improve practical relevance. On the algorithmic side, we propose SDDC-YOLO, a YOLOv8-based defect detection framework that integrates frequency-domain feature modeling and dense multi-scale feature fusion. A frequency-domain attention mechanism is introduced to suppress interference from textured backgrounds, while a densely connected feature fusion neck enhances the perception of defects across multiple spatial scales. Together, these components form an effective and application-oriented framework for woven fabric defect detection.