Crack detection in civil infrastructure, including roads, bridges, and buildings, is crucial for maintaining structural safety and functionality. Traditional manual inspection methods are time-consuming and prone to e...
详细信息
Crack detection in civil infrastructure, including roads, bridges, and buildings, is crucial for maintaining structural safety and functionality. Traditional manual inspection methods are time-consuming and prone to errors, highlighting the need for automated solutions. This study evaluates state-of-the-art computer vision techniques for automatically detecting cracks in both asphalt and concrete surfaces from 2013 to 2024. The study assesses the effectiveness and limitations of image processing, traditional machine learning, and deep learning methods for crack detection. A comparative analysis of commonly used models is presented, utilizing public datasets: SDNET2018, CCIC, and BCD for concrete images, and AigleRN, CFD, CRACK500, and GAPs for asphalt images. Based on the comparison results, advanced deep learning models such as YOLOv5 and U-Net have demonstrated superior performance in crack detection for both asphalt and concrete structures, significantly outperforming traditional methods. For concrete crack detection, YOLOv5l exhibited exceptional performance on the SDNET2018 dataset, achieving a precision of 97.7%, recall of 96.7%, and a mAP@.5 of 99.3%, with a rapid inference time of 1.1 ms, making it highly suitable for real-time applications. For asphalt crack detection, U-Net achieved outstanding results, particularly on the GAPs dataset, with a near-perfect precision of 99.53%, and on the CFD dataset, with a precision of 92.54% and an F1-score of 89.90%. The study also highlights public concrete and asphalt datasets, providing details on methodology, including the number of images, image sizes, and noted noise factors. Additionally, it discusses the impact of data source variability on crack detection methods, showcasing the applications, strengths, and limitations of multi-sensor fusion techniques. Finally, unresolved challenges such as imbalanced datasets, high inference times, and complex network architectures are identified, with suggestions for future
暂无评论