Training-Free Logical and Structural Anomaly Detection via Calibrated Fusion

To address the need for simultaneously handling structural and logical anomalies in industrial anomaly detection, researchers proposed a calibration fusion method that integrates heterogeneous anomaly cues without training. This method utilizes statistical information from normal images to align anomaly features from different sources and directly fuses complementary frozen features within a unified framework, without additional training or component-level supervision. Experiments showed that this method achieved image-level AUROC rates of 95.9% for logical anomalies and 89.0% for structural anomalies on the MVTec-LOCO dataset, with an average of 92.5%, outperforming all existing untrained detectors and matching performance with methods that require network training or component labeling. Additionally, its structural anomaly variant achieved an image-level AUROC of 99.1% on the MVTec-AD dataset, comparable to PatchCore, demonstrating that this calibration mechanism is versatile and can be applied beyond logical anomaly detection.