CiC-NET: a real-time semantic segmentation network for dam surface crack detection

Crack detection is vital for maintaining hydraulic engineering infrastructure. However, achieving a balance between real-time processing and high precision in semantic segmentation models presents a significant challenge, especially given the intricate details of cracks and complex backgrounds. To tackle this issue, this paper proposes a real-time, high-precision crack segmentation model. Initially, a four-branch feature extraction structure is devised to capture the edge details of cracks, with a focus on enhancing segmentation accuracy. Subsequently, an image pyramid is constructed at the input end to feed small-scale samples into high-dimensional feature extraction branches, thereby reducing computational costs and improving segmentation speed. Finally, an effective feature fusion module is designed for the feature extraction structure to capture sufficient crack features and achieve precise crack localization. Extensive experiments validate the superior performance of the proposed method, with Pixel Accuracy, Recall, Intersection over Union, and F1 score reaching 68.45 \(\%\) $\%$ , 67.04 \(\%\) $\%$ , 51.22 \(\%\) $\%$ , and 67.74 \(\%\) $\%$ , respectively, while maintaining a modest parameter count of 13.12M. This model provides a reliable and efficient solution for the health monitoring and maintenance of hydraulic engineering.