Unifying transformer and convolution for dam crack detection

Abstract

Cracks are a serious disease that threatens the safety of a hydraulic dam. However, detecting dam cracks in time is still a challenging task. It was found that the existing methods fail to detect elongated cracks, leading to discontinuous crack detection results. Moreover, the detection accuracy is severely affected due to the diversity and complexity of cracks and backgrounds. To address these problems, we propose a pixel-level crack detection network by unifying the transformer and CNN models (UTCD-Net). Specifically, the transformer model is designed to extract the global context features for detecting long-range cracks and removing distracting background information, the CNN model is employed to extract local feature information for the detection of thin cracks, and the final segmentation result is refined by the attention fusion module. In comparison with current mainstream methods, the proposed method can capture both local and global features of cracks, which are helpful to detect thin and long cracks. The method obtained promising results on our self-built dam dataset in terms of the Intersection of Union and F1 score (IoU:64.08%, F1:49.43%). Moreover, the experimental results on the other three public crack datasets demonstrate that the proposed method is flexible for adapting to different scenarios. The self-built dam crack dataset provides a challenging benchmark for future research.