摘要
许多图像去雾算法采用无监督学习以提高在真实场景中的泛化能力,然而,现有的无监督去雾算法大多依赖循环生成对抗网络,通过简单的去雾和加雾循环来实现,对生成图像缺乏有效约束,导致去雾效果不理想。此外,这些方法在特征提取时未能充分考虑雾的物理特性,使得恢复的图像纹理细节信息丢失严重。为了解决以上问题,文中提出一种基于物理解耦与自适应特征提取的无监督图像去雾网络(UPDA-Net)。首先,设计物理解耦网络(PDN),通过大气散射模型建立物理约束机制,分离和估计雾化图像中的大气光值与传输图,更准确地还原图像的光照信息,提升去雾过程的物理一致性及图像清晰度;其次,引入自适应特征提取模块(AFEM),结合大气光与传输图的物理特性,在特征空间中应用大气散射模型对相关特征进行近似,合成清晰图像的潜在物理特征,以增强网络模型的特征表达能力,改善恢复图像的细节和纹理质量。实验结果表明,在多个公开真实图像去雾数据集上,该算法在评价指标和视觉效果上均优于近年来的六种主流去雾算法。
Unsupervised learning is adopted to improve generalization in real scenarios in many image dehazing algorithms.However,the existing unsupervised dehazing methods mostly rely on the CycleGAN(cycle-consistent generative adversarial network)framework and realize dehazing by simple cycles of dehazing and re-hazing,so it lacks effective constraints on the generated images and results in suboptimal performance.Moreover,these methods fail to sufficiently consider the physical properties of haze in the feature extraction,which causes a severe loss of texture details.Therefore,an unsupervised image dehazing network based on physical decoupling and adaptive feature extraction is proposed,and the network model is named UPDA-Net.Specifically,a physical decoupling network(PDN)is designed to separate and estimate the atmospheric light and transmission map in hazy images by establishing physical constraints based on the atmospheric scattering model.This design allows for more accurate restoration of image illumination,and enhances the physical consistency and clarity of the dehazing process.Furthermore,an adaptive feature extraction module(AFEM)is developed to integrate the physical characteristics of atmospheric light and transmission maps.By applying the atmospheric scattering model in the feature space,this module approximates relevant characteristics and fuses the potential physical features of clear images,so as to enhance the model's feature representation capability and improve the detail and texture quality of the restored images.Experimental results demonstrate that the proposed method outperforms six mainstream dehazing algorithms in terms of objective evaluation and visual quality on several public real image dehazing datasets.
作者
闫在爽
贺鹏
YAN Zaishuang;HE Peng(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,Yichang 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China)
出处
《现代电子技术》
北大核心
2025年第17期77-84,共8页
Modern Electronics Technique
关键词
图像去雾
大气散射模型
物理解耦
参数估计
自适应特征提取
无监督学习
循环生成对抗网络
图像恢复
image dehazing
atmospheric scattering model
physical decoupling
parameter estimation
adaptive feature extraction
unsupervised learning
CycleGAN
image restoration
作者简介
闫在爽(1997-),男,湖北襄阳人,硕士研究生,研究方向为图像处理、深度学习;通讯作者:贺鹏(1965-),男,湖北宜昌人,硕士研究生,教授,研究方向为区块链技术、图像处理。
