A review of fog image generation algorithms based on deep learning
基于深度学习的雾图生成算法综述

摘要

雾图生成技术旨在将清晰图像合成为具有真实雾霾视觉效果的图像，是计算机视觉领域的重要研究方向。该技术不仅是增强图像去雾模型训练效果、提升视觉系统在复杂环境下鲁棒性的关键技术手段，还在自动驾驶仿真测试、影视特效制作、数字艺术创作等实际场景中展现出广泛应用价值。传统雾图生成方法高度依赖大气散射模型的参数估计，受限于模型简化性与参数设计主观性，生成雾图的真实感与场景适应性均存在明显不足。近年来，深度学习技术的快速发展，尤其是生成对抗网络与扩散模型的突破性应用，为雾图生成技术注入了新的发展动力。本文系统梳理了基于深度学习的雾图生成算法研究进展，将其划分为物理模型引导型、端到端生成型与扩散模型驱动型三大类别，深入剖析各类方法的核心原理、技术优势及适用局限。最后，结合当前研究现状，总结该领域面临的关键挑战，并从技术融合、功能拓展、性能优化等角度展望未来研究方向，为后续相关研究提供参考。

Abstract

Fog image generation technology aims to synthesize clear images into images with realistic foggy visual effects and is an important research direction in the field of computer vision. This technology is not only a key means to enhance the training effectiveness of image dehazing models and improve the robustness of visual systems in complex environments, but also demonstrates broad application value in practical scenarios such as autonomous driving simulation tests, film special effects production, and digital art creation. Traditional fog image generation methods rely heavily on parameter estimation of atmospheric scattering models. Due to the simplification of the models and the subjectivity in parameter design, both the realism and scene adaptability of the generated fog images are significantly limited. In recent years, the rapid development of deep learning technology, especially the breakthrough applications of generative adversarial networks and diffusion models, has injected new momentum into fog image generation technology. This paper systematically reviews the research progress of deep learning-based fog image generation algorithms, categorizing them into three major types: physics model-guided, end-to-end generation, and diffusion model-driven. It provides an in-depth analysis of the core principles, technical advantages, and applicable limitations of each type. Finally, in combination with the current research status, the paper summarizes the key challenges in this field and, from the perspectives of technology integration, functional expansion, and performance optimization, explores future research directions, providing a reference for subsequent related studies.