Scalable and Accurate Crack Segmentation for Infrastructure Health Monitoring Using EfficientNetB0 and a Context-Aware Attention Mechanism

In response to the growing need for accurate and automated infrastructure assessment, this study introduces a novel crack segmentation framework that combines an EfficientNetB0 encoder with an enhanced U-Net decoder augmented by contextual attention mechanisms. Designed to operate effectively in complex and variable environments, the model demonstrates superior performance in accurately detecting and segmenting cracks. To enhance generalization and mitigate overfitting, the architecture incorporates dropout layers, while synthetic data generation addresses the challenge of limited annotated data sets. The total data set contains 11,292 images, with 1,695 images allocated for testing and 9,603 images for training. Experimental evaluations reveal a training accuracy of 98.39% and a Dice coefficient of 0.7028, surpassing traditional segmentation approaches in both precision and robustness. In addition to segmentation, the system incorporates crack line detection, further improving its precision and utility. The model’s capacity to adapt to real-world conditions underscores its potential for deployment in critical infrastructure monitoring. By enabling early and reliable crack detection, the proposed system contributes to proactive maintenance strategies, ultimately reducing operational costs and enhancing public safety. These results highlight the effectiveness and practicality of the proposed method for efficient structural health monitoring applications.