Two-Stage Nighttime Desnowing Diffusion Model Based on Pseudo-Scenario Reconstruction

Nighttime image restoration under adverse weather conditions is more challenging than daytime due to snow's refractive and scattering properties and uneven illumination from active artificial light sources. Firstly, to address the absence of comprehensive benchmark datasets, we proposed a method for pseudo-scenario reconstruction to get high-quality images that can be used for nighttime snowing image restoration network training. The method separates the global illumination distribution and introduces a dynamic adaptive coefficient for the snow mask, then reconstructs the image details to restore the original lighting and object reflectance. Building upon this, we further considered the light color effect to generate realistic nighttime snowfall scenes that contain the influence of brightness and color. Besides, we use the conditional diffusion model to guide the nighttime desnowing model as a strong benchmark, effectively eliminating snow noise influenced by localized light. Experimental results indicate the superior performance of our method over current state-of-the-art methods in nocturnal desnowing efficacy in real-world scenarios and having good generalization.