SoftPatch is an unsupervised anomaly detection method designed for computer vision and visual inspection tasks that operate on training datasets containing noisy or mislabeled images. Building upon memory-based feature matching, the approach analyzes image data at the local patch level to differentiate normal patterns from anomalous noise. It employs noise discriminators to calculate outlier scores for patch-level feature representations, filtering out corrupted patches before constructing a representative coreset in a memory bank. By incorporating these outlier scores into the decision process, the method softens the anomaly detection boundaries, mitigating overconfidence and enabling robust defect identification and localization in practical, noise-prone environments.
