Abstract: Road surface crack detection is a critical task in road maintenance and safety management. Cracks in road surfaces are often the early indicators of larger structural issues, and if not detected and repaired in time, they can lead to more severe deterioration and increased maintenance costs. Effective and timely crack detection is essential to prolong road lifespan and ensure the safety of road users. This paper introduces CrackNet, an advanced crack detection model built upon the YOLOv9 architecture, which integrates a fusion attention module and task space disentanglement to enhance the accuracy and efficiency of road surface crack detection. Traditional methods often struggle with the complex and irregular nature of road cracks, as well as the challenge of distinguishing cracks from their backgrounds. CrackNet overcomes these challenges by leveraging an attention mechanism that highlights relevant features in both the channel and spatial dimensions while separating the tasks of classification and regression. This approach significantly reduces false negatives and improves localization accuracy. The effectiveness of CrackNet is validated through comparative analysis with other segmentation models, including Unet, SOLO v2, Mask R-CNN, and Deeplab v3+. CrackNet consistently outperforms these models in terms of F1 and Jaccard coefficients. This study highlights the critical role of accurate crack detection in minimizing maintenance costs and enhancing road safety.

Quanwu Li and Shaopeng Duan, “Road Surface Crack Detection Based on Improved YOLOv9 Image Processing” International Journal of Advanced Computer Science and Applications(IJACSA), 15(11), 2024. http://dx.doi.org/10.14569/IJACSA.2024.0151156

@article{Li2024,
title = {Road Surface Crack Detection Based on Improved YOLOv9 Image Processing},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2024.0151156},
url = {http://dx.doi.org/10.14569/IJACSA.2024.0151156},
year = {2024},
publisher = {The Science and Information Organization},
volume = {15},
number = {11},
author = {Quanwu Li and Shaopeng Duan}
}