摘要
在多种恶劣天气条件下,捕获的图像经常受到图像退化的困扰,这对图像质量和进一步的视觉分析带来了显著挑战。现有的多场景修复模型通常侧重于区分不同天气间的特定特征,借助大型复杂网络处理天气间的扰动,这导致模型参数规模较大。针对这一问题,该文提出了一种基于多尺度傅里叶变换双分支网络(MFTNet)的多天气图像修复方法。傅里叶校正子网通过结合快速傅里叶变换和Unet网络提取全局频率信息,缩小不同天气间的噪声差异并有效降低了模型参数量。多尺度细化子网通过归一化特征分布和动态提取多尺度特征,进一步学习丰富的特征表示,有效消除天气间的相似物理退化信息。MFTNet以较小的参数成本提升了图像的最终重建质量,其网络参数量相比All-in-One降低了86.32%,相比TransWeather降低了84.18%。在RESIDE、Raindrop、Outdoor-Rain、Snow100K和Rain1400等数据集上的实验结果显示,该方法在SSIM和PSNR等评价指标上均得到明显的提升。
Captured images often suffer from image degradation under multiple adverse weather conditions,which poses significant challenges to image quality and further visual analysis.Existing multi-scene restoration models usually focus on distinguishing specific features between different weather conditions and rely on large complex networks to handle the disturbances between weather conditions,which results in a large model parameter scale.To address this problem,we propose a multi-weather image restoration method based on multi-scale Fourier transform two-branch network(MFTNet).The Fourier correction subnetwork extracts global frequency information by combining fast Fourier transform and the Unet network,which reduces the noise differences between different weather conditions and effectively reduces the number of model parameters.The multi-scale refinement subnetwork further learns rich feature representations by normalizing feature distributions and dynamically extracting multi-scale features to effectively eliminate similar physical degradation information between weather.MFTNet improves the final reconstruction quality of the image with a small parameter cost.Its network parameters are reduced by 86.32%compared to All-in-One and 84.18%compared to TransWeather.Experimental results on RESIDE,Raindrop,Outdoor-Rain,Snow100K,and Rain1400 datasets show that the proposed method has been significantly improved in evaluation metrics such as SSIM and PSNR.
作者
马佳慧
张鸿
李富豪
MA Jia-hui;ZHANG Hong;LI Fu-hao(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System(Wuhan University of Science and Technology),Wuhan 430065,China)
出处
《计算机技术与发展》
2025年第8期198-205,共8页
Computer Technology and Development
基金
国家重点研发计划(2020AAA0108503)。
关键词
图像修复
恶劣天气
全局频率信息
多尺度特征表示
快速傅里叶变换
image restoration
adverse weather
global frequency information
multi-scale feature representation
fast Fourier transform
