TRAC-AD: Transformation-Response-Aware Correction for Anomaly Detection

Unsupervised visual anomaly detection supports industrial quality inspection using defect-free reference images. Nearest-neighbor detectors can nevertheless assign high anomaly scores to normal regions when illumination or other acquisition conditions change. We propose Transformation-Response-Aware Correction for Anomaly Detection (TRAC-AD), a residual-scoring method built on a frozen DINOv2 backbone. Spatially paired feature differences between normal source images and their photometrically transformed views define a low-dimensional response subspace. During inference, TRAC-AD retrieves normal references in the original fused feature space, removes the response-subspace component of the deepest-layer matching residual, and combines its norm with unchanged scores from the other layers. On AeBAD-S, RobustAD-derived PCB and MetalParts settings, and MVTec AD 2 TESTpub, the mean gains in normalized area under the per-region overlap curve up to a 5% false-positive rate ([email protected]) are 3.20, 1.14, and 0.49 percentage points over the matched base pipeline, respectively. Original-resolution TESTpub evaluation yields a mean [email protected] gain of 0.90 percentage points. These results support using paired photometric responses to improve localization at low false-positive rates under acquisition changes.