Synthetic Data-Driven U-Net Semantic Segmentation for Automated Defect Detection

Abstract

This study presents the development, training, and comparative evaluation of a U-Net-based convolutional neural network for pixel-level semantic segmentation of surface defects on structural components in nuclear reactor environments. Four defect classes are considered: cracks, corrosion, voids/pores, and background. Due to the limited availability of real annotated inspection data, the model is trained exclusively on a synthetically generated dataset consisting of 2000 images. The central research question addresses the impact of data augmentation on generalization under synthetic-to-real domain shift. Two models with identical architectures are evaluated: Model A (with augmentation) and Model B (without augmentation). Despite the lower validation mIoU (0.9353 vs. 0.9814), Model A demonstrates significantly better generalization on real images, avoiding the systematic overclassification of corrosion characteristic of Model B (average 15.5× area overestimation). These findings indicate that augmentation plays a critical role in mitigating domain shift and improving model reliability in cross-domain settings. The study further identifies a failure mode termed overconfident domain overfitting, where models trained on synthetic data achieve high validation performance but produce systematically biased and overconfident predictions on real data. Overall, the results suggest that synthetic data, when combined with appropriate augmentation strategies, can be a viable basis for training segmentation models under controlled conditions, although further validation on diverse architectures and real annotated datasets is required.