Fine-grained multimodal anomaly detection via dual-stream representations and tri-evidence adaptive fusion
Measurement Science and Technology · 19.06.2026 · 37 (25) · 256111
DOI: 10.1088/1361-6501/ae758d ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 23.06.2026. Verfasst von Zhuolun Li und 2 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben.
Abstract (Originalsprache)
Multimodal anomaly detection is important for real world applications such as public opinion monitoring and data security. However, most existing approaches primarily produce coarse decisions at the sample or segment level and pay limited attention to fine-grained anomalies, especially those caused by noise or short duration local perturbations. In addition, modality reliability may change abruptly over time due to occlusion or acquisition failures. This often causes sample level detectors to yield unstable and difficult to interpret anomaly scores when modality quality becomes imbalanced. To address these issues, we propose FC-UDNet, a fine-grained cross-modal detection network with uncertainty and evidence aware fusion. FC-UDNet adopts a dual-stream complementary encoder that learns representations from a local sliding window attention stream and a global similarity graph stream. We further introduce an anomaly aware graph consistency regularization that encourages smoothness among reliable time steps while reducing contamination from suspicious nodes, thereby stabilizing normal structural patterns. Moreover, FC-UDNet incorporates a tri-evidence driven reliability modality gating mechanism. The gate jointly models reconstruction residuals, predicted variances, and cross-modal discrepancies as complementary reliability evidence, enabling the model to assign lower weights to unreliable modalities and provide interpretable modality specific cues. The resulting evidence is aggregated into a unified anomaly score at each time-step for accurate localization of complex multimodal anomalies. Experiments on the public CMU-MOSI and CMU-MOSEI datasets show that FC-UDNet consistently outperforms representative baselines. Ablation studies and visualization analyses further validate the contribution of each component and the robustness under modality degradation.
Zitieren
APA
Li, Z., Gu, R., & Li, C. (2026). Fine-grained multimodal anomaly detection via dual-stream representations and tri-evidence adaptive fusion. Measurement Science and Technology, 37(25), 256111. https://doi.org/10.1088/1361-6501/ae758d
Vancouver
Li Z, Gu R, Li C. Fine-grained multimodal anomaly detection via dual-stream representations and tri-evidence adaptive fusion. Measurement Science and Technology. 2026;37(25):256111. doi:10.1088/1361-6501/ae758d
BibTeX
@article{Li2026,
author = {Li, Zhuolun and Gu, Ruichun and Li, Chaofeng},
title = {Fine-grained multimodal anomaly detection via dual-stream representations and tri-evidence adaptive fusion},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {25},
pages = {256111},
doi = {10.1088/1361-6501/ae758d}
}
EUVIMED