StyleGAN3-driven instance-aware dynamic fusion augmentation with lightweight YOLO-MD for defect detection on microstructured surfaces

Microstructured surfaces are critical in modern manufacturing, yet online defect detection still faces challenges such as limited samples, redundant networks, data imbalance, and complex backgrounds. To address this, this study proposes a StyleGAN3-based Instance-Aware Dynamic Fusion Augmentation (SIDFA) method and a lightweight YOLO-MD model. SIDFA can extract defect features from small, multidimensionally imbalanced datasets and adaptively fuse them into real image backgrounds. YOLO-MD reduces model complexity through L1 regularization and structured pruning, while knowledge distillation is employed to maintain detection performance. On the self-built Microstructure Injection-molded Defect Dataset (MIDD), YOLO-MD (SIDFA) achieves an mAP@0.5 of 82.541% with only 1.813 M parameters. Compared with YOLO26n (SIDFA) trained on the same augmented dataset, YOLO-MD (SIDFA) slightly improves detection accuracy while reducing the number of parameters by 23.66% and increasing the frame rate (FPS) by 11.58%. Compared with the non-augmented YOLO26n, mAP@0.5 is improved by 10.882%. Under the same conditions, YOLO-MD (SIDFA) achieves a better balance between detection accuracy and model complexity than mainstream models such as YOLOv8s (79.265%, 11.127 M) and YOLOv10n (79.162%, 2.266 M). These results indicate that the proposed method has potential for real-time industrial deployment and automated quality inspection of microstructured surfaces. The MIDD dataset are publicly available at: https://github.com/EZ4JJ/MIDD-YOLO-MD .