2026 Volume 7 Issue 3 Pages 807-820
In the inspection of outdoor structures, changes in imaging conditions can cause domain shifts between surveys, potentially degrading anomaly detection performance. In this study, we apply survey-contained zero-shot anomaly detection, in which model construction and inference are performed using only unlabeled data acquired in the target survey, to river revetment inspection and evaluate its effectiveness. In this setting, the training and test data are identical, thereby avoiding domain shifts between surveys. However, because the unlabeled data contain a mixture of normal and anomalous samples, constructing a model while suppressing the influence of anomalies is a key challenge. In the experiments, conventional PatchCore and the survey-contained versions of PatchCore, SoftPatch, SubspaceAD, and MuSc were compared under the same conditions. Among the survey-contained methods, SoftPatch, which incorporates a noise removal mechanism based on LOF, achieved the highest performance, with mean pixel-level AUPRO values of 84.9% for nunozumi and 84.7% for tanizumi. These results demonstrate the applicability of survey-contained zero-shot anomaly detection to damage screening using the river revetment dataset.