Developing reliable segmentation procedures for distinguishing material phases in asphalt mixtures remains challenging due to their highly heterogeneous microstructure. This challenge becomes even more pronounced in mixtures containing high contents of reclaimed asphalt pavement (RAP), where overlapping X-ray attenuation responses limit the ability of conventional image-analysis techniques to consistently discriminate between neighbouring phases in X-ray computed tomography (XCT) images. To address this limitation, this study presents a multiclass deep-learning segmentation framework based on a U-Net architecture for the simultaneous identification of six classes within XCT images of high-RAP asphalt mixtures: cracks and voids, mastic, RAP fine asphalt mixture (RAP-FAM), RAP coarse aggregates, virgin steel-slag aggregates, and background. The proposed pipeline integrates image pre-processing, semi-automatic and manual annotation, data augmentation, patch-based training, weighted loss optimisation, and three-dimensional volume reconstruction. The trained model achieved high segmentation performance, with an average Intersection over Union (IoU) of 0.890 considering all classes and 0.868 when excluding the background class, together with average precision, recall, and F1-score values of 0.908, 0.954, and 0.929, respectively. The framework proved effective in handling severe class imbalance, enabling the crack-and-void class to achieve a recall of 0.991 and an F1-score of 0.922. Three-dimensional renderings revealed elongated low-density features preferentially located along virgin aggregate boundaries and enabled the visualisation of spatial relationships between virgin aggregates, RAP-derived phases, and damage-related features. Although cracks and voids were intentionally grouped into a single class, the reconstructed morphologies provided qualitative evidence consistent with crack development. The proposed methodology offers a scalable and reproducible solution for phase-resolved XCT analysis and provides a robust basis for investigating microstructural damage mechanisms in highly heterogeneous recycled asphalt mixtures.
Work package 2
.png){kind=logo}
