A machine learning‐based multi‐scale morphological approach for crack segmentation

Abstract To achieve quantitative analysis of cracks in different regions for C f /SiC composites, this study proposes an automated, multi‐scale crack segmentation and quantitative analysis framework for composites with complex backgrounds. This framework is based on the ResNet50‐Unet model, which achieves accurate segmentation of fibers and matrix. In addition, adaptive grayscale thresholding combined with multi‐scale morphological pore elimination effectively suppresses the influence of pseudo‐cracks (defects or pores) while extracting cracks. Moreover, a KD‐Tree‐based crack connection algorithm module repairs micro‐crack fractures, significantly improving crack connectivity. The results demonstrate that the machine learning model achieves a segmentation accuracy of 97%. Following the pore elimination process, the crack line density, area density, and average width were significantly reduced from 18.8 mm/mm 2 , 7.2 mm 2 /cm 2 , and 6.8 µm to 1.1 mm/mm 2 , 0.2 mm 2 /cm 2 , and 2.2 µm, respectively. Subsequent to the crack connection procedure, the crack line density, area density, and total crack length increased from 1.1 mm/mm 2 , 0.23 mm 2 /cm 2 , and 1.8 cm to 1.6 mm/mm 2 , 0.31 mm 2 /cm 2 , and 2.1 cm, respectively. These findings validate the effectiveness of the proposed crack extraction framework in enabling precise crack quantification and offer a novel methodology for crack analysis in ceramic matrix composites.