FM-Dehaze: Fine-grained modeling of non-uniform haze degradation for unpaired image dehazing
Shuaibin Fan, Wenchao Yan, Minglong Xue, Palaiahnakote Shivakumara, Tong Lu · Optics & Laser Technology · 2026
Images captured under hazy atmospheric conditions typically exhibit significant contrast degradation and detail loss, which severely compromises the perceptual performance of outdoor vision systems. To address this challenge, this paper focuses on resolving the fundamental difficulty of achieving an optimal balance between modeling spatially non-uniform haze distributions and maintaining global consistency in complex scenes. We propose an image dehazing method of fine-grained modeling for non-uniform haze degradation. First, at the feature representation level, based on the Kolmogorov-Arnold theorem, we design a decoupled-collaborative dual-branch channel representation mechanism that explicitly decomposes and implicitly aggregates nonlinear mapping relationships. This enhances the network’s ability to perceive spatially heterogeneous haze distributions and generates restoration results that better align with human visual perception. Second, at the modeling paradigm level, we break from traditional methods’ reliance on feature extraction modules and atmospheric scattering models, reformulating haze degradation as a continuous signal mapping problem in implicit function space. By leveraging implicit neural representations, we efficiently parameterize degradation intensity at arbitrary spatial coordinates using compact network parameters, thereby mitigating redundant information propagation. Third, at the representation enhancement level, we integrate a dense residual enhancement module that, through cascaded multi-scale residual connections, progressively disentangles and eliminates haze artifacts to generate high-fidelity dehazed results. Experimental results demonstrate that extensive evaluations on multiple synthetic benchmarks and real-world scenarios validate the effectiveness of our approach. The project code will be available at https://github.com/Fan-pixel/FM-Dehaze .