Normal Tail Suppression for Low-False-Positive RGB–3D Industrial Anomaly Localization

Abstract

RGB images and 3D point clouds provide complementary appearance and geometric cues for industrial anomaly localization, yet a small number of high-scoring normal regions can still produce spurious defect candidates under stringent false-positive constraints. We propose Normal Tail Suppression (NTS), a training-stage regularizer that selectively penalizes the upper tail of normal fused scores while leaving the deployed inference architecture unchanged. By concentrating optimization on the normal responses most relevant to low-false-positive-rate (FPR) operation, NTS avoids uniformly shrinking the normal-score distribution. On MVTec 3D Anomaly Detection (MVTec 3D-AD), NTS improves mean AUPRO@1% by +0.352 percentage points over the matched control and maintains a positive mean trend across five paired seeds. Sensitivity, control-ablation, threshold-resolution, and score-distribution analyses support the intended tail-focused behavior, while a protocol-frozen second fusion configuration provides further positive evidence. These results establish NTS as a simple, deployment-compatible mechanism for improving low-FPR RGB–3D anomaly localization within the evaluated configurations.