Beyond architecture: Hyperparameter optimization for YOLOv8m in multi-class textile defect detection

• Presented a configuration-first hyperparameter optimization approach for YOLOv8m applied to multi-class textile defect detection. • Achieved 63.2% mAP@0.5 with real-time processing throughput, surpassing more complex architectural models. • Demonstrated robustness through multi-seed reproducibility tests with low variance and 95% confidence intervals. • Provided comprehensive ablation studies clarifying the rationale for loss weights, classification thresholds, and validation settings. • Benchmarked against state-of-the-art models including EfficientDet, transformers, and anchor-free methods with consistent evaluation settings. Fabric defect detection in industrial textile manufacturing remains a challenging task due to intricate patterns and subtle defect features. Although existing methods employ sophisticated architectural innovations, they often introduce significant computational overhead and complexity, limiting industrial applicability. This study presents a configuration-centric approach for multi-class textile defect detection, optimizing the hyperparameters of the baseline YOLOv8m model without modifying its architecture. Experiments on the Tianchi Fabric Dataset with 22 defect classes involved systematic tuning of augmentation strategies, confidence and IoU thresholds, and loss function parameters through a rigorous multi-phase search procedure. The optimized YOLOv8m achieves 65.2% mean Average Precision at IoU 0.5 (mAP@0.5), surpassing several recent state-of-the-art models: the YOLOX-CATD framework (54.6% mAP@0.5), JDCBL transformer-based method (39.3%), AMFF multi-scale fusion network (39.1%), and CRFB approach trained on fewer classes (61.5%). Notably, it marginally outperforms FD-YOLOv5 (65.1% mAP@0.5) on a more extensive class set and exceeds RT-DETR-L transformer detector (64.1%) while requiring 24% fewer GFLOPs and exhibiting superior inference speed (17.4 FPS on NVIDIA GeForce RTX 3060 12GB). Ablation studies validate the significant impact of hyperparameter choices on performance, and saliency-based visualization confirms accurate defect localization despite interference from fabric motifs. These findings emphasize the efficacy of hyperparameter optimization as a scalable alternative to architectural complexity, offering a balanced trade-off between accuracy, speed, and deployment readiness in real-world production lines. The novelty of this study advances fabric defect inspection by providing a reproducible, efficient, and interpretable framework, underscoring disciplined configuration as a pivotal factor for state-of-the-art performance.