基于全局-局部先验和纹理细节关注的图像修复  

Image Inpainting Based on Global-Local Prior and Texture Details

作　　者：徐祺津 叶海良 曹飞龙 梁吉业[3] XU Qijin;YE Hailiang;CAO Feilong;LIANG Jiye(Department of Applied Mathematics,College of Sciences,China Jiliang University,Hangzhou 310018;School of Mathematical Sciences,Zhejiang Normal University,Jinhua 321004;Department of Artificial Intelligence,School of Computer and Information Technology(School of Big Data),Shanxi University,Taiyuan 030006)

机构地区：[1]中国计量大学理学院应用数学系,杭州310018 [2]浙江师范大学数学科学学院,金华321004 [3]山西大学计算机与信息技术学院(大数据学院)人工智能系,太原030006

出　　处：《模式识别与人工智能》2025年第2期101-115,共15页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.62176244)资助。

摘　　要：图像修复旨在利用周围信息填充图像中的缺失区域,然而现有基于先验的方法大多难以兼顾全局语义一致性和局部纹理细节.因此,文中提出基于全局-局部先验和纹理细节关注的图像修复方法,结合小波卷积与傅里叶卷积,构造小波-傅里叶卷积块,增强局部特征和全局特征的交互.在此基础上,提出全局-局部学习式先验,通过一个由小波-傅里叶卷积块构成的先验提取器,同时学习全局先验和局部先验.该先验提取器作用于受损图像和完整图像,分别得到受损先验和监督先验.在修复阶段,受损图像和学习的先验分别输入两个结构相似的修复分支.这两个分支均由小波-傅里叶卷积构成,能同时提取和融合全局特征与局部特征.最后,合并两个分支的输出,生成具有一致语义内容和清晰局部细节的图像.此外,构造高感受野风格损失,从语义层面提升图像风格一致性.实验表明,文中方法在多个数据集上均性能较优.Image inpainting is intended to fill in missing regions of an image using surrounding information.However,existing prior-based methods often struggle to balance global semantic consistency and local texture details.In this paper,a method for image inpainting based on global-local prior and texture details is proposed.Wavelet-Fourier convolution blocks are constructed by combining wavelet convolution and Fourier convolution to enhance the interaction between local and global features.Based on the above,a global-local learning-based prior is presented.A prior extractor composed of wavelet-Fourier convolution blocks is designed to simultaneously learn global and local priors.The prior extractor is applied to both damaged and complete images to obtain damaged priors and supervised priors.During the repair phase,the damaged image and the learned priors are input into two structurally similar repair branches.Both branches are constructed with wavelet-Fourier convolutions and can simultaneously extract and fuse global and local features.Finally,the outputs of the two branches are merged to generate the image with consistent semantic content and clear local details.Additionally,a high receptive field style loss is introduced to improve image style consistency at the semantic level.Experimental results show that the proposed method outperforms existing methods on multiple datasets.

关 键 词：图像修复 学习式先验 小波变换 全局-局部特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]