Abstract
Developing reliable segmentation procedures to observe fatigue crack propagation in asphalt mixtures is challenging due to material heterogeneity containing, particularly when they contain high reclaimed asphalt pavement contents. Specifically, when images are X-ray computed tomograms (XCT), overlapping X-ray attenuation responses by the distinct material phases make conventional image analysis methods fail to consistently distinguish cracks from surrounding phases. In this study, a multiclass deep learning segmentation pipeline based on a U-Net architecture is presented for the simultaneous identification, within XCT, of the background and five material phases (crack and voids, mastic, PM-RAP fine asphalt mixtures (FAM), PM-RAP coarse aggregates, and virgin aggregates). The pipeline integrates pre-processing, semi-automatic and manual labelling, patch-based training, and volumetric rendering, enabling a robust phase discrimination across large XCT. The trained model achieved high segmentation performance, with a mean Intersection over Union (mIoU) of 0.89, accurately differentiating aggregates, mastic regions, and cracks. Three-dimensional renders of the segmented phases revealed preferential crack initiation along virgin aggregate–binder interfaces and allowed detailed assessment of the interaction between PM-RAP components. This pipeline provides a scalable and reproducible methodology for phase-contextualised crack analysis, offering new insights into fatigue damage mechanisms in heterogeneous asphalt systems and supporting the development of more durable recycled pavement materials.

This publication has 36 references indexed in Scilit: