Beyond Static Fusion: Instance-Adaptive Confidence Reallocation for Multimodal Emotion Recognition
IEEE Signal Processing Letters · 2026 · 33 · 2909-2913
DOI: 10.1109/lsp.2026.3708418 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Kunming Wu und 3 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben.
Abstract (Originalsprache)
Multimodal Emotion Recognition (MER) suffers from a fundamental mismatch between static fusion strategies and the dynamic nature of emotional expression: the informativeness of each modality varies significantly across emotional states, contexts, and individuals. Conventional methods assume fixed modality contributions during inference, leading to suboptimal performance when certain modalities carry weak or ambiguous signals. To bridge this gap, we propose the Multimodal Router Weighted Fusion (MRWF) framework, a dynamic fusion approach that couples an Adaptive Feature Weighting (AFW) module with an information-theoretic confidence constraint. AFW performs hierarchical feature extraction through multi-path transformations, enabling fine-grained, multi-path modality representation. Grounded in information theory, the confidence constraint regularizes this extraction by enforcing a structured lower bound on joint modality contribution, promoting stable adaptive weight reallocation under modality conflict or incompleteness. This co-design enables instance-tailored fusion with negligible architectural overhead. Extensive experiments demonstrate that MRWF effectively addresses sample-level modality inconsistency and mitigates performance degradation under missing modality scenarios, yielding more stable training dynamics and robust fusion across diverse MER benchmarks.
Zitieren
APA
Wu, K., Kou, Y., Liu, H., & Xu, Z. (2026). Beyond Static Fusion: Instance-Adaptive Confidence Reallocation for Multimodal Emotion Recognition. IEEE Signal Processing Letters, 33, 2909-2913. https://doi.org/10.1109/lsp.2026.3708418
Vancouver
Wu K, Kou Y, Liu H, Xu Z. Beyond Static Fusion: Instance-Adaptive Confidence Reallocation for Multimodal Emotion Recognition. IEEE Signal Processing Letters. 2026;33:2909-2913. doi:10.1109/lsp.2026.3708418
BibTeX
@article{Wu2026,
author = {Wu, Kunming and Kou, Yunfeng and Liu, Hong and Xu, Zijing},
title = {Beyond Static Fusion: Instance-Adaptive Confidence Reallocation for Multimodal Emotion Recognition},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2909--2913},
doi = {10.1109/lsp.2026.3708418}
}
EUVIMED