ENHANCED YOLOV8 MODEL FOR REAL-TIME FABRIC DEFECT DETECTION IN TEXTILE INDUSTRY

Qarshiyev B. E. & Kholliyev S. R.


Annotatsiya

In the textile industry, automated defect detection is crucial for maintaining product quality and reducing production costs. Traditional methods often struggle with real-time performance and accuracy on high-resolution images. This paper proposes an improved YOLOv8-based model, termed FabricDefect-YOLO (FD-YOLO), incorporating three key enhancements: a Reparameterized Grouped C2f (RG-C2f) module for efficiency, Dynamic Sampling (DySample) upsampling for better detail preservation, and an Adaptive Inner-Wise Intersection over Union (AI-WIoU) loss function for robust bounding box regression. Evaluated on the TILDA and Tianchi datasets, FD-YOLO achieves a mean Average Precision (mAP@0.5) of 82.4% on TILDA and 91.5% on Tianchi, outperforming the baseline YOLOv8 by 6.4% and 1.5%, respectively, while maintaining real-time inference speeds of over 120 FPS on standard hardware. These improvements make FD-YOLO suitable for industrial deployment in resource-constrained environments.


Adabiyotlar

  1. Kumar, A. (2008). Computer-vision-based fabric defect detection: A survey. IEEE TIE.
  2. Ngan, H. Y. (2011). Automated fabric defect detection—A review. Image Vis. Comput.
  3. Jing, J. (2019). Fabric defect detection using deep CNN. IEEE Access.
  4. Ronneberger, O. (2015). U-Net: Convolutional networks for biomedical image segmentation. MICCAI.
  5. Ren, S. (2015). Faster R-CNN: Towards real-time object detection. NeurIPS.
  6. Liu, W. (2016). SSD: Single shot multibox detector. ECCV.
  7. Redmon, J. (2016). You only look once: Unified, real-time object detection. CVPR.
  8. Jocher, G. (2023). Ultralytics YOLOv8.
  9. Wang, C. Y. (2023). YOLOv7: Trainable bag-of-freebies. arXiv.
  10. Li, Y. (2022). Improved YOLOv5 for fabric defects. Sensors.
  11. Ding, X. (2021). RepVGG: Making VGG-style conv architectures great again. CVPR.
  12. Liu, J. (2023). DySample: Dynamic positional encoding for efficient upsampling. arXiv.
  13. Tong, Z. (2023). Wise-IoU: Bounding box regression loss. arXiv.
  14. Zheng, Z. (2023). Inner-IoU for better overlap. IEEE TPAMI.