U-shaped high-resolution network for semantic segmentation of concrete cracks

ABSTRACT The safety and service life of concrete structures heavily rely on continuous evaluation and maintenance. Traditional manual inspection methods are not only labor-intensive and costly but also pose safety risks. This study presents Crack-UHRNet, an innovative semantic segmentation model designed for high-precision detection of concrete cracks using convolutional neural networks. By adopting a U-shaped High-Resolution Network (U-HRNet) as its backbone and incorporating the context representation of OCRNet, this model significantly enhances segmentation accuracy. Trained on diverse datasets of concrete surface cracks across various environments, Crack-UHRNet shows remarkable generalization capability and superior performance, particularly in challenging backgrounds and noisy conditions. Additionally, this study developed a fracture quantification workflow to extract geometric features—area measurement, length via skeletonization, width estimation through distance transformation, and orientation analysis through linear fitting. This study not only advances automated crack detection technology, but also provides a robust tool for effective infrastructure maintenance and safety assessment.