PDCFuse: Prior-driven degradation-controlled fusion for nighttime infrared-visible image fusion

Infrared and visible image fusion under nighttime conditions aims to generate a fused image that jointly preserves thermal target saliency and visible structural details. The main difficulty is that the visible modality is frequently degraded by low illumination, noise, blur, and local glare, which weakens structural reliability and makes fixed fusion strategies unreliable across different scenes. Existing methods usually perform static cross-modal aggregation and seldom explicitly model visible-image degradation to adapt the fusion and reconstruction process. In this study, we propose PDCFuse, a prior-driven degradation-controlled framework for nighttime infrared-visible image fusion. The proposed method first estimates a three-dimensional visible degradation state and then uses it to jointly guide visible feature correction, infrared information injection, detail modulation, and decoder-side visible skip reconstruction. To make the degradation controller learnable and stable, a night-oriented training scheme with visible-only synthetic degradation, masked degradation supervision, relative clean-anchor regularization, and clean-degraded joint validation is further designed. In addition, a degradation-target saliency and naturalness (DTSN) objective is introduced to enhance weak thermal targets while suppressing background disturbance and bright-region overflow. The experiments combine broad benchmark comparisons with a dedicated evaluation against nighttime-oriented and degradation-aware methods on fixed nighttime subsets. The results show favorable overall fusion performance, stronger robustness to nighttime degradation, a more balanced target-background representation, practical inference efficiency, and benefits for downstream object detection.