Answers are grounded in this paper and cite their sources.
Accurate detection and characterization of air voids in asphalt core slices are essential for evaluating pavement mesostructure. However, existing deep-learning segmentation methods are rarely evaluated under concurrent pixel-level class imbalance, limited annotated images, and small irregular voids, and the propagation of segmentation errors into geometric and spatial descriptors is seldom examined. This study develops a U-Net++-based framework for void detection and distribution analysis under imbalanced small-sample conditions. A Combo Loss combining weighted binary cross-entropy and Dice loss is used to alleviate pixel-level class imbalance. U-Net++, U-Net, Attention U-Net, YOLOv8, and YOLOv11 are compared under identical settings; the YOLO models are evaluated separately as an object-level detection paradigm. On a single 52/13 image-level split, U-Net++ achieved the highest validation F1 (0.8168), precision (0.8101), recall (0.8258), and IoU (0.6904) among the segmentation models. Because the same 13 validation images were used for early stopping, threshold selection, model selection, and reporting, these values are validation-tuned and exploratory; no independent test set or external validation was performed. Predicted masks were qualitatively examined for void size, morphology, and spatial arrangement; no quantitative symmetry or morphology validation was conducted. The framework provides a preliminary basis for automated asphalt-core void characterization under imbalanced small-sample conditions.
No takes yet. Share an insight, caveat, or question.
Zheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: