基于混合编码和掩膜空间调制的图像补全算法  

Image Inpainting Algorithm Based on Hybrid Encoding and Mask Space Modulation

作　　者：冼进[1] 徐小茹[2] 冼允廷 冼楚华[1] XIAN Jin;XU Xiaoru;XIAN Yunting;XIAN Chuhua(School of Computer Science and Engineering,South China University of Technology,Guangzhou 510006,Guangdong,China;School of Law and Humanities,Jiangsu Vocational College of Finance and Economics,Huaian 223003,Jiangsu,China)

机构地区：[1]华南理工大学计算机科学与工程学院,广东广州510006 [2]江苏财经职业技术学院法律与人文艺术学院,江苏淮安223003

出　　处：《华南理工大学学报(自然科学版)》2025年第3期31-39,共9页Journal of South China University of Technology(Natural Science Edition)

基　　金：广东省重点领域研发计划项目(2022B0101070001);华南理工大学中央高校基本科研业务费专项资金资助项目(Z2kjD2220210)。

摘　　要：图像补全是指对图像缺失区域进行内容填充,是计算机视觉和图像处理研究的重要问题之一。当前图像补全算法的研究已经取得了很大的进展,但如果图像中的场景比较复杂且存在大面积的缺失区域时,现有算法由于缺乏有效的网络结构来捕捉图像的长距离依赖和高级语义信息,仍然较难生成高质量的完整图像。针对大范围缺失的图像补全问题,为扩大图像补全网络有限的感受野,有效地获取图像可见区域的全局信息,同时充分利用图像可见区域的有效信息,该文提出了一种基于混合编码和掩膜空间调制的图像补全算法。该算法首先通过混合编码网络对图像可见区域进行局部和全局信息的特征提取;然后采用掩膜空间调制模块,根据缺失面积的大小动态调整在生成缺失区域时的多样性;最后基于StyleGAN2的方法生成完整图像。实验结果表明,该文提出的算法能够有效地处理大范围缺失的图像,可生成具有多样性的高质量图像,并且能应用在视觉显著性模型的数据增强上。Image inpainting refers to the process of filling in missing regions of an image with plausible content,which is one of the significant issues in the fields of computer vision and image processing research.Current research on image inpainting algorithms has made substantial progress.However,when dealing with complex images with large missing areas,existing algorithms still face challenges in generating high-quality complete images due to the lack of effective network structures to capture long-range dependencies and high-level semantic information in the images.To address the issue of large-scale missing image inpainting,this paper proposed an image inpainting algorithm based on hybrid encoding and mask spatial modulation.The aim is to expand the limited receptive field of image inpainting networks,effectively obtain global information from the visible regions of the image,and fully utilize the effective information from the visible regions.Firstly,a hybrid encoding network was used to extract local and global information features from the visible regions of the image.Then,a mask spatial modulation module dynamically adjusted the diversity in generating missing regions based on the size of the missing area.Finally,a method based on StyleGAN2 was used to generate complete images.Experimental results show that the proposed algorithm can effectively handle images with large-scale missing areas,generating high-quality images with diversity,and can be applied to data augmentation in visual saliency models.

关 键 词：图像补全 图像增强 混合编码 掩膜空间调制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]