Low-Light Consistency Learning for Degradation-Suppressed Infrared and Visible Image Fusion

Abstract

Infrared - visible image fusion aims to integrate thermal structure with visible appearance information. Under low illumination, however, photometrically unstable visible responses may be mistaken for useful scene details and transferred to the fused image. To address this problem, a two-view low-light consistency mechanism is introduced before cross-modal fusion. Brightness-attenuated and contrast-reduced views are constructed from the visible luminance and processed by a shared encoder. A feature-level consistency constraint reduces the discrepancy between their representations, which are then averaged and integrated with infrared features through base - detail fusion. Experiments on LLVIP show that the proposed method achieves the highest $\boldsymbol{Q}_{\boldsymbol{A} \boldsymbol{B} \boldsymbol{F}}$ among the compared approaches while maintaining strong source-image correlation. Ablation results confirm the contribution of the consistency constraint and show that the two photometric views provide complementary improvements. Without target-domain fine-tuning, the model also obtains leading $\boldsymbol{Q}_{\boldsymbol{A} \boldsymbol{B} \boldsymbol{F}}$ and CC results on RoadScene. These results indicate that the proposed design helps limit the influence of photometrically unstable low-light responses while preserving source-edge information and source-image correlation.