摘要
为了提高神经网络模型从全局到局部的特征提取能力,增强雾图复原过程中的泛化性能,提出一种结合宽型提取和径向增强的二阶段图像去雾算法.算法通过两个阶段对有雾降质图像进行清晰化复原,在特征提取阶段,结合Transformer的全局关注能力与可变形卷积的局部感知能力,提出由双尺度Transformer模块和可变形卷积模块组成三分支宽型特征提取结构,实现对雾图特征有效感知与提取;在特征增强阶段,采用密集残差块组成的径向增强网络,依次联合由浅到深的图像特征进行拼接,实现特征的进一步增强.实验结果表明:所提出的网络模型在合成数据集和真实图像复原过程中均表现出色,对不同浓度的雾霾去除效果明显,复原图像主观恢复效果自然,且具有较好的泛化能力.
To address the need for improving the neural network model's ability to extract features from global to local contexts and enhancing the generalization performance in haze image restoration,a two-stage image dehazing algorithm combining wide-scale extraction and radial enhancement was proposed.The proposed algorithm performed clear restoration on degraded images with haze in two stages.In the feature extraction stage,a three-branch wide-scale feature extraction structure was proposed,composed of a dual-scale Transformer module and a deformable convolution module,and this structure combined the global attention capability of the Transformer with the local perception capability of deformable convolutions to effectively perceive and extract features from hazy images.In the feature enhancement stage,a radial enhancement network consisting of dense residual blocks was utilized,and this network progressively concatenated image features from shallow to deep levels,further enhancing the extracted features.Experimental results show that the proposed network model performs exceptionally well in both synthetic datasets and real image restoration processes,showing significant haze removal effects for different haze concentrations,and the restored images exhibit natural subjective recovery and possess good generalization capabilities.
作者
杨燕
陈阳
张浩文
YANG Yan;CHEN Yang;ZHANG Haowen(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;State Grid Gansu Electric Power Company Wuwei Power Supply Company,Wuwei 733000,Gansu China)
出处
《华中科技大学学报(自然科学版)》
北大核心
2025年第5期164-170,共7页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61561030,62063014)
甘肃省高等学校产业支撑计划资助项目(2021CYZC-04)
兰州交通大学教改项目(JG201928)
甘肃省教育厅优秀研究生“创新之星”资助项目(2023CXZX-547)。
