Abstract
Infrared and visible images provide complementary cues for scene interpre-tation: infrared measurements highlight thermally salient objects, whereasvisible measurements preserve fine structures and textures. Their fusionis challenging because adaptive modality selection can amplify one sourcebefore the two modalities have formed a compatible representation. Wepropose ConRef-Fusion, a two-stage latent fusion network that first learnsa shared-decoder-compatible representation and then performs local refine-ment around this representation. In the first stage, modality-specific en-coders and a shared decoder are optimized with an equal-weight latent an-chor. In the second stage, the encoders are fixed and a lightweight consensusresidual adapter predicts a bounded correction from the latent discrepancybetween the two modalities. This design preserves a stable cross-modal ref-erence while allowing scene-dependent enhancement of thermal targets andvisible details. Experiments on the TNO, M3FD, RoadScene, and MSRSbenchmarks show that the proposed method provides competitive entropy,complementary-information, fidelity, and edge-preservation scores. Objectdetection experiments on M3FD further indicate that the fused images retaincues useful for high-level perception. Ablation and latent-decoding analyses clarify how the consensus anchor and residual refinement contribute to theobserved performance

This publication has 20 references indexed in Scilit: