From Detection to Segmentation: Adapting YOLO26 for Surface Crack Delineation

Abstract

Surface crack delineation plays an important role in infrastructure inspection because accurate pixel-level mapping of cracks can support condition assessment, maintenance planning, and automated monitoring of roads and other civil surfaces. However, recent crack segmentation methods often depend on heavy network designs or prompt-based foundation models. It also remains unclear how a lightweight detection-oriented model behaves across different data organizations and unseen images. This study therefore investigates whether YOLO26m-seg, the medium segmentation variant of the recent YOLO26 family, can be adapted into an effective and practical pipeline for surface crack delineation. The proposed workflow has four main steps. First, binary crack masks are converted into polygon labels. Second, the model is fine-tuned on crop-based, full-image, and merged training splits, with preprocessing and augmentation applied during training. Third, the predicted instance masks are merged into binary crack maps. Fourth, the outputs are evaluated with pixel-level overlap and classification measures on both in-domain and unseen test data. The findings show that the adapted model produces competitive crack masks across the examined protocols. On CRACK500 it achieved 67.20% Precision, 80.01% Recall, and 72.12% Dice, and it maintained stable behavior in fold-based validation. It also transferred meaningfully to the unseen DeepCrack benchmark, reaching 69.54% Precision, 78.64% Recall, and 71.24% Dice without task-specific redesign. The main contributions are a mask-to-polygon adaptation pipeline for YOLO26m-seg, an evaluation across several data protocols with zero-shot transfer to an unseen benchmark, and a comparison with published methods reported as context. Analyses of threshold sensitivity, ablation settings, and computational efficiency support them. These results indicate that an efficient detection-oriented segmentation pipeline can provide a practical alternative to heavier or prompt-dependent crack segmentation systems. Overall, the study supports the conclusion that moving from detection to segmentation within a lightweight deployment-friendly framework is both feasible and useful for automated surface inspection.