Fracture analysis is essential for identifying the causes of metal product failures and preventing failure recurrence. This study focuses on automating the segmentation of fracture surfaces, addressing challenges such as overexposure and underexposure clipping in scanning electron microscopy (SEM) images. We propose a deep learning model enhanced with a novel data augmentation technique called clipping augmentation. This technique artificially introduces clipping effects, such as overexposure and underexposure, into the images to improve segmentation accuracy. Our experiments utilized 1000 SEM images of fracture surfaces, labeled at the pixel level by experts. The dataset was divided into training, validation, and test sets. Model architecture evaluations revealed that the Attention U-Net with ResNet50 fine-tuning provided the highest intersection over union (IoU) scores, achieving a remarkable score of 0.933 on test data. Optimal preprocessing included contrast-limited adaptive histogram equalization and pseudo-coloring, significantly enhancing segmentation performance. Clipping augmentation, by tuning the maximum artificial clipping size and the number of artificial clippings per image, markedly improved the model's robustness against actual overexposure and underexposure artifacts. Combined with standard data augmentation techniques, this method significantly improved the IoU, demonstrating the efficacy of our approach in handling complex SEM images. Our results indicate that the proposed method can reliably segment fracture surfaces in SEM images, even under adverse conditions, paving the way for more efficient and accurate fracture analysis automation.
