Evaluation of Training Data Selection Strategies for Automated Concrete Surface Crack Detection

Abstract

In deep-learning-based concrete crack detection, degraded generalization on unseen bridges due to domain shift remains a major practical challenge. This study compares two robust training-data selection strategies: target-specific sampling, which prioritizes images with textures similar to the inference target, and cluster-balanced sampling, which ensures coverage of diverse background patterns through feature clustering. Bridge-wise cross-validation on real bridge-deck datasets revealed that target-specific sampling overfits to particular background features, leading to both missed detections (false negatives) and over-detections (false positives). In contrast, cluster-balanced sampling reduces reliance on specific textures by learning diverse background patterns, thereby suppressing missed detections and maintaining generalization performance on unseen bridges. These findings provide practical guidelines for constructing training datasets that enhance the reliability of AI models for infrastructure inspection.