A Lightweight Fusion Method for Few-Shot Industrial Anomaly Detection Using Color Images and Point Clouds

Abstract

Few-shot industrial anomaly detection is difficult because only a small set of normal training samples is available, while defects may appear as appearance changes, structural changes, or both. In this setting, combining color-image evidence with point-cloud geometry can improve robustness, but direct multimodal fusion is often unstable and difficult to calibrate. To address this problem, this paper proposes a lightweight fusion method that preserves the original dual-branch backbone and changes only the fusion and scoring stages. The method introduces a retrieval-guided sample-level re-ranking strategy and a low-weight feature-level auxiliary fusion branch so that normal-reference calibration and feature-assisted scoring can be combined without replacing the base encoders. On the benchmark dataset used in this study, the proposed auxiliary fusion setting achieves an image-level receiver operating characteristic area under the curve of $0.923~\pm ~0.005$ across three random seeds in a unified-code comparison, compared with $0.903~\pm ~0.006$ for the reproduced re-ranking baseline. The paired improvement is positive for all three seeds, with an average gain of $0.019~\pm ~0.002$ . The three-seed average is also slightly higher than the 0.919 point-cloud-plus-color-image value reported in the reference study. The method also improves pixel-level receiver operating characteristic area under the curve, F1 score, and area under the precision-recall curve in both branches over the corresponding single-modality results reported in the reference study. A supplementary image-level check on the MVTec three-dimensional anomaly-detection benchmark shows a small positive but non-significant trend, so the cross-dataset result is treated as supporting evidence rather than as a strong generalization claim. These results indicate that a carefully calibrated lightweight fusion stage can strengthen few-shot multimodal anomaly detection while retaining branch-wise localization behavior.