TN-YOLO: A YOLOv8-Based Algorithm for Extremely Small Defect Detection

Extremely small defect detection poses critical challenges in industrial quality inspection, particularly for defects characterized by micron-scale dimensions, irregular geometries, and low signal-to-noise ratios in complex backgrounds. While existing YOLO-series models suffer from feature degradation, redundant sampling, and noise amplification in such scenarios, this paper proposes TN-YOLO with three dedicated innovations: 1) The Adaptive Feature Integration (AFI) module employing learnable weights to maintain cross-level defect features; 2) The TNet neck optimizing feature fusion paths to preserve shallow spatial details; 3) The Normalized Channel Attention Mechanism (NCAM) leveraging dynamic channel-wise calibration to synergistically enhance defect features while suppressing background noise. Experimental validation on both a custom industrial wheel-hub DK dataset and the public SeaDronesSeev2 benchmark (specifically designed for extremely small objects) demonstrates TN-YOLO's superiority: 8.3% AP50 improvement over YOLOv8 on SeaDronesSeev2 while maintaining 36.6% mAP@[.50:.95] on DK-surpassing all compared models. The framework exhibits 19.7-20.4% relative improvement in mAP@[.50:.95], proving exceptional robustness in low-SNR environments. These advancements provide novel insights for small defect detection in industrial applications.