Proposal for an improved segmentation method for fracture surfaces with overexposure and underexposure clipping

Fracture surface analysis is vital for clarifying failure mechanisms of metallic products and for preventing recurrence. This study focuses on automating the segmentation of scanning electron microscopy (SEM) fracture images, where challenges often arise from overexposure and underexposure. To address this, we apply a deep learning model with a new data augmentation strategy termed clipping augmentation. By artificially inserting clipping effects, the model gains robustness against imaging artifacts. A dataset of 1000 expert-labeled fracture images was divided into training, validation, and test sets. Among the evaluated models, Attention U-Net with ResNet50 fine-tuning achieved the best performance, with an intersection over union (IoU) of 0.933 on test data. Preprocessing with adaptive histogram equalization and pseudo-coloring further improved accuracy. These results demonstrate that the method can provide reliable fracture segmentation under difficult conditions. The approach is also expected to be valuable in the field of light metals, such as aluminum and magnesium alloys, where efficient fracture analysis supports safety and durability.