Fabric defect detection using fine-tuned Yolo-12

Area of Science:

• Textile Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Automatic defect identification in textiles is crucial for quality control, reducing waste and costs.
• Traditional manual inspection is time-consuming and prone to errors.
• Deep learning offers a promising solution for efficient and accurate fabric defect detection.

Purpose of the Study:

• To develop a robust and computationally efficient deep learning model for identifying a wide spectrum of fabric defects.
• To ensure the model performs effectively on both plain and printed textiles.
• To validate the model's performance using both local and publicly available datasets.

Main Methods:

• The study proposes UniDefectNet-Omni, a fine-tuned and optimized YOLOv12 model.
• Key techniques include high determination feature learning, heterogeneous defect representation, and adaptive augmentation.
• The model was trained on a local dataset from Chenab Textiles and validated on Tildav2, DPFD-DET, and ZJU-Leaper datasets.

Main Results:

• UniDefectNet-Omni achieved high performance across datasets, with mAP scores ranging from 85.1% to 93.6%.
• On the Chenab dataset, it reported 85.1% mAP, 84.5% precision, and 81.7% recall for seven defect categories.
• The model demonstrated strong performance on public datasets, including 86.7% mAP on TILDAv2 and 93.6% mAP on DPFD-DET.

Conclusions:

• UniDefectNet-Omni is a lightweight, efficient, and robust framework for fabric defect detection.
• The model demonstrates excellent generalization capabilities across diverse fabric types and defect categories.
• This approach significantly enhances automated fabric quality inspection systems.