Crack detection and semantic segmentation in concrete structures using an improved deep learning-based framework

Abstract Early recognition of cracks in concrete structures is vital to avert progressive deterioration, unexpected failures, and costly maintenance interventions. Conventional manual inspection is subjective, labor-intensive, and difficult to apply over large-scale infrastructure. In addition, many existing automated approaches still suffer from reduced robustness under shadows, surface noise, and complex concrete textures. To tackle these issues, this exploration develops an enhanced DL-based framework for automatic crack detection and pixel-level segmentation on concrete surfaces. The recommended scheme was trained and assessed on a database of 69 concrete surface images, and its performance was assessed using standard statistical metrics. On the held-out test database, the recommended scheme achieved a precision of 98.08%, a recall of 90.68%, and an F1-score of 91.38%, indicating a good balance between correctly identifying crack pixels and limiting false alarms, while requiring on average 1.2 s to process each image. From a practical standpoint, the proposed approach provides an effective tool to support automated condition assessment of concrete bridges, tunnels, pavements, and building elements, enabling objective, repeatable, and rapid crack evaluation. Therefore, this research contributes to improving the reliability of structural health monitoring systems and highlights the urgency of adopting intelligent vision-based inspection methods for aging concrete infrastructure.