Deep Learning-Based Microcrack Segmentation for Damage Evolution Assessment in SiCf/SiC Composites

The inherent brittleness of SiC f /SiC composite claddings causes stress concentration and microcrack initiation at micro-defects. Early detection of these cracks under micro-strain is difficult due to their small scale, irregular morphology, and background noise. To address this, an enhanced U-Net model integrated with a novel dual-source attentive block module (DSABM) is proposed. The DSABM effectively fuses multi-scale features and improves discrimination between cracks and background interference. Combined with strategic data augmentation and dynamic dataset splitting, the method demonstrates robust performance. On a dedicated cladding crack dataset, it achieves an F1-score of 86.84% and a precision of 89.21% in detecting cracks larger than 10 μm, while processing images at 165 frames per second. The model shows strong generalization in complex scenarios, confirming its utility as a reliable, high-speed automated tool for the inspection and safety assessment of nuclear ceramic cladding. • In terms of data management, this work addresses the challenges of limited sample size and imbalanced crack scale distribution. A dynamic augmentation pipeline coupled with a morphology-aware stratified splitting algorithm is introduced, which substantially improving model generalization, robustness, and evaluation reliability. • Architecturally, a dual-source attentive block module (DSABM) is incorporated into the U-Net framework, which enhances discriminative feature extraction for SiCf/SiC cladding micro-cracks while suppressing irrelevant background noise, effectively addressing the insufficient accuracy of conventional approaches in complex imaging conditions. • A dynamic threshold optimization algorithm (DTOA) is developed to adaptively determine the optimal binarization threshold based on validation performance. Integrated with a morphological post-processing pipeline, the algorithm can effectively suppresses noise, restores crack continuity, and enables highly precise crack identification.