Fabric defect detection using fine-tuned Yolo-12

Defect identification is critical for ensuring the reliability and price of fabrics. Defective fabrics result in significant waste and losses. Automatic defect identification using the use of deep learning is a faster and more efficient way to analyze fabric quality, replacing human inspection. Furthermore, both plain and printed textiles are produced concurrently in enterprises; hence, one design should be effective in identifying faults in both types of fabric. As a result, a strong deep learning algorithm must be trained to detect defects within fabric datasets produced during manufacturing with excellent performance and cheap computing costs. This study incorporates a local dataset collected from Chenab Textiles and validated with three publicly available datasets, such as Tildav2, DPFD-DET, and ZJU-Leaper. The experiment provides a comprehensive and diversified range of defective images. To identify textile defects successfully, the suggested approach, universal and optimized YOLOv12, named universal defect detect network (UniDefectNet-Omni) for robust identification of a wide spectrum of fabric defects with multi and diverse types of fabric using YOLOv12 by fine-tuning and optimizing training, integrating high determination feature learning, heterogeneous defect representation, and adaptive augmentation. As a consequence, UniDefectNet-Omni is a lightweight, computationally efficient, and robust framework across varied fabrics.The Chenab textile dataset mean Average Precision (mAP) is 85.1%, precision is 84.5%, and recall is 81.7% over seven separate fabric defect categories. The proposed fine-tuned YOLOv12 outperformed on validated datasets, such as the TILDAv2 dataset, having a mean Average Precision (mAP) score 86.7%, precision about 83.7%, in addition recall about 83.6% over four separate fabric defect categories. DPFD-DET with a mean Average Precision (mAP) score 93.6%, precision is 92.2%, and recall is 87.8% over four separate fabric defect categories. ZJU-Leaper with groups 1, 2, 3, and 4 having a mean Average Precision (mAP) score 93%, precision about 78.1%, in addition recall about 90.3% over twelve separate fabric defect categories.