Beyond architecture: Hyperparameter optimization for YOLOv8m in multi-class textile defect detection
Year: 2025
Type: article
Abstract: Fabric defect detection presents unique challenges due to faint, fine-grained anomalies embedded in repetitive textile patterns. While recent studies enhance accuracy through architectural innovations such as attention modules or transformer heads, these solutions often increase complexity and hinder deployment. This study proposes a configuration-first approach: we retain the base YOLOv8m architecture and systematically optimize its hyperparameters for the Tianchi Fabric Dataset, a 22-class industrial benchmark. By disabling MixUp and Mosaic, calibrating confidence and NMS thresholds, applying modest Copy-Paste augmentation, and tuning loss weights, our model achieves 63.2% [email protected], surpassing five state-of-the-art YOLO variants including CRFB, JDCBL, and AMFF. Notably, it outperforms more complex models like RT-DETR-L on [email protected]:0.95 while requiring 24% less compute, with real-time throughput on a Tesla T4. A detailed ablation study isolates the contribution of each hyperparameter. Moreover, D-RISE saliency maps confirm that the tuned model accurately localizes defects under challenging texture interference. This work demonstrates that strategic configuration alone without architectural change can rival transformer-class accuracy, supporting a scalable path for real-world deployment. Our findings advocate a shift from capacity-driven to configuration-aware optimization in industrial vision tasks.
Source: Results in Engineering
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,940
OpenAlex ID: W4416443791