MIFC-Net: Joint compression and fusion for infrared-visible imaging with state-space models and invertible feature decoupling

Joint compression and fusion of paired infrared-visible images under transmission constraints is a key challenge for edge imaging systems. Existing methods often incur considerable overhead in global context modeling, do not facilitate differentiated probabilistic modeling of multimodal latent representations, and waste bits on sensor noise. We propose MIFC-Net, a joint compression and fusion framework that combines state-space modeling with invertible latent reparameterization. The state-space encoder uses 2D selective scanning to capture long-range cross-modal dependencies; INNFlow organizes the fused latent representation into Base and Detail paths, which are modeled by separate probability models; the noise-aware objective and PQiM modulation constrain reconstructed content and the spatial scale of the Detail path. Experiments on multiple public datasets show that MIFC-Net maintains good fusion quality under bandwidth constraints while retaining information useful for downstream object detection.