Adaptive dual-domain-guided semantics diffusion model for image defogging
Shaohui Jin, Yanxin Zhang, Ziqin Xu, Zhengguang Qin, Xiaoheng Jiang, Hao Liu, Mingliang Xu · Optics & Laser Technology · 2025
Seeing through dense fog is a significant challenge in computer vision for attenuation of the target information, leading to differences in visual content. Image defogging works based on diffusion models still suffer from limitations such as loss of content and color information, as well as constraints in terms of long inference times. This paper introduces an adaptive dual-domain-guided semantics diffusion model for image defogging, which integrates the robust anti-interference capabilities of laser range-gated imaging with the exceptional feature learning abilities of diffusion models. Specifically, a deterministic data adaptive enhancer (DDAE) is utilized to analyze features in both the frequency and spatial domains, employing kernel-level features (such as blur intensity) extracted from blurry images. Subsequently, an enhanced-discrepancy-stochastic refinement (EDSR) strategy is employed to enable the conditional denoising diffusion model (CDDM) to perceive the non-ideal regions between foggy images and high-quality enhanced images, ultimately achieving high-fidelity image reconstruction. Extensive experiments have demonstrated that our method exhibits exceptional fog perception capabilities and outstanding defogging performance on indoor and outdoor laser range-gated imaging data with varying optical thicknesses (ranging from 0.5 to 3.0) as well as on color data.