Abstract: Crack detection on concrete surfaces is essential for maintaining the structural integrity of aging infrastructure. However, supervised deep learning models often suffer from performance degradation when the training and test data distributions differ. This paper addresses covariate shift caused by differences in image resolution—termed Multi-Resolution Adaptation (MRA)—by introducing a one-step framework for crack segmentation. The proposed method jointly trains a U-Net–based segmentation model and a lightweight importance-estimation model that assigns weights to training images according to their estimated likelihood of appearing in the test domain. The training in this method is done in one step, reducing effects of underfitting inherent in two-step approaches and only requires unlabeled images in target domains. Class Activation Mapping (CAM) is employed to visualize the regions contributing to importance estimation, revealing that resolution differences are captured mainly through background surface texture rather than crack structures. Experiments conducted on a public dataset with artificially generated multi-resolution images demonstrate that the proposed approach improves Intersection over Union (IoU), particularly when low-resolution images constitute as little as 1% of the training data. These results confirm that one-step covariate shift adaptation effectively enhances model robustness to multi-resolution domain shifts while requiring only unlabeled target domain images.

Keywords: Crack segmentation; covariate shift; multi-resolution; importance weighting.