LRIF-Net: Decision-Preserving Lightweight Recurrent Interaction Fusion for Multimodal Crisis Information Classification

Abstract

Emergency-information analysts need to organize image-text reports into humanitarian categories before reports can be verified and prioritized. This study asks whether fine-grained evidence refinement can be added to an established classifier without replacing its decision path or greatly enlarging the task-specific update. We propose LRIF-Net, a decision-preserving recurrent residual fusion framework that retains a complete frozen global-local classifier and adds a compact parameter-tied correction branch. The branch is trained and evaluated with three predefined, synthetic evidence-degradation conditions—missing image, missing text, and different-class image mismatch—so the resulting evidence concerns these controlled protocols rather than every form of real-world data corruption. On the canonical 7314-pair CrisisMMD Task 2 split, the fixed three-seed ensemble achieves 91.35% Accuracy, 88.88% Macro F1, and 91.32% Weighted F1. The residual stage updates 517,509 parameters, 88.76% fewer than matched standard cross-attention, while reducing all retained non-CLIP parameters by 59.17%. LRIF-Net provides competitive performance under the matched evaluation protocol, but paired comparisons do not establish predictive superiority and its Balanced Accuracy and Macro F1 are lower than those of the retained direct classifier. Ablations attribute the controlled degradation gains mainly to the learning objective. The study contributes the coordinated design and evaluation of a compact residual refinement mechanism for crisis-information triage, with predictive and robustness claims limited to the reported evidence.