| 摘要: |
| 在雾天条件下采集复杂交通场景中的数据，往往面临较高的成本投入与安全风险。为此，提出了一种融合Stable Diffusion与光学模型的雾气图像生成方法，在保持原图像结构特征的基础上，能够合成高质量的雾天图像。该方法首先利用双目图像与相机参数估算场景深度，优化透射率图后结合随机森林回归模型估计大气光照强度生成雾图。随后，借助ControlNet生成结构化深度图，引导Stable Diffusion的生成过程，并通过大语言模型的低秩适配机制微调U型卷积神经网络与语言-图像对比预训练模型，以提升模型的雾天表达能力。最终，融合两种模拟结果生成高质量雾天图像。量化与可视化实验结果表明，所提出的融合算法在保持原始图像特征的同时，相较于Optical Model-based Fog Simulation在轻度、中度和重度雾条件下的真实性值分别提升了48.71%、217.92%和502.16%。此外，生成的仿真雾天图像显著增强了YOLOv11n在不同雾浓度场景中的目标检测性能。 |
| 关键词: 智能交通 目标检测 雾气图像 Stable Diffusion 光学模型 视觉感知 |
| DOI：10.20079/j.issn.1001-893x.250530002 |
|
| 基金项目:国家自然科学基金资助项目（52072259）；天津市自然科学基金（24JCZDJC00490） |
|
| Foggy Image Generation and Visual Perception Methods for Intelligent Transportation |
| LIU Xinchao,ZHU Yueying,LI Yuwei,LIN Ye |
| (College of Mechanical Engineering，Tianjin University of Science and Technology，Tianjin 300222，China) |
| Abstract: |
| Collecting data from complex traffic scenarios under foggy conditions often incurs high costs and poses significant safety risks.To this end，a fog image generation method integrating Stable Diffusion with an optical model is proposed，which is capable of synthesizing high-quality foggy images while preserving the structural characteristics of the original images.Specifically，scene depth is first estimated using stereo images and camera parameters.The transmission map is then optimized，and the atmospheric light intensity is estimated via a random forest regression model to generate preliminary fog images.Subsequently，ControlNet is employed to produce structured depth maps that guide the Stable Diffusion process，while a Low-Rank Adaptation mechanism is used to fine-tune the U-shaped convolutional neural network and the Contrastive Language-Image Pre-training model，thereby enhancing the model’s capability to represent foggy scenes.Finally，the results of the two simulation branches are fused to generate high-fidelity foggy images.Quantitative and visual experimental results demonstrate that the proposed hybrid method not only preserves the original image features but also improves the authenticity value by 48.71%，217.92%，and 502.16% under light，moderate，and heavy fog conditions，respectively，compared with traditional Optical Model-based Fog Simulation.Moreover，the generated foggy images significantly enhance the performance of YOLOv11n for object detection under varying fog densities. |
| Key words: intelligent transportation object detection foggy image Stable Diffusion optical model visual perception |
《电讯技术》编辑部 