Skip to content
Open access

Gki-net: Gaussian splatting and KAN-based implicit neural representation for physically grounded dehazing

Oct 2026 · Signal, Image and Video Processing · Vol 20 · 0 citations · 15 references

Abstract

Haze causes severe image degradation in visual sensing and obscures key physical cues. We propose GKI-Net, a physics-embedded dehazing framework that integrates Gaussian splatting with a convolutional Kolmogorov–Arnold network. Instead of relying on black-box mappings, we design a Hybrid Implicit Physics Decoder (HIPD). Using multi-scale implicit neural representations, the decoder explicitly decouples the transmission map and atmospheric light in the atmospheric scattering model, thereby preserving physical consistency in the restored images. During feature encoding, we design a Gaussian feature splatting module that uses deformable Gaussian kernels to adaptively aggregate long-range spatial context. We further introduce a convolutional KAN layer (ConvKAN) to perform efficient high-order nonlinear feature transformations with a small number of parameters. Experimental results show that GKI-Net achieves leading or competitive performance in quantitative metrics and visual perception, and attains a good balance between computational efficiency and restoration quality, thereby effectively improving the environmental adaptability of sensing systems under complex weather conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.