Hyperparameter Optimization Outperforms Architectural Complexity: A Systematic Study on Neural Network Design for Industrial Fabric Defect Detection

Industrial fabric defect detection requires high accuracy across heterogeneous defect types while maintaining edge-compatible computational efficiency. This study challenges the assumption that architectural complexity improves performance by demonstrating that systematic hyperparameter optimization achieves superior accuracy-efficiency trade-offs. Through controlled ablation experiments on the Tianchi dataset (3,985 images, 22 defect classes), we evaluate ten YOLOv12m configurations: baseline, eight architecture-modified variants (CBAM, SE, P2 multi-scale, RepNCSPELAN4, ATSM), and one hyperparameter-optimized variant. All eight architectural modifications degrade performance (8.2%–29.8% mAP@0.5 reduction), while hyperparameter-optimized YOLOv12m achieves 64.8% mAP@0.5 with 20.2M parameters, a 12.9% improvement over baseline while maintaining identical computational cost, 100% pretrained weight transfer, and deployment simplicity. This configuration surpasses 12 of 14 published methods while being 22% lighter than comparable architectures. Root cause analysis reveals three failure modes: optimization complexity mismatch, data efficiency limitations with insufficient supervision for complex attention mechanisms, and fabric-specific inductive bias mismatch. For fabric defect detection with limited training data and severe class imbalance, systematic hyperparameter optimization outperforms architectural modifications while preserving deployment feasibility. These results suggest optimization-first strategies warrant exploration in textile inspection before pursuing architectural complexity.