基于Transformer结构的图像修复算法研究综述  

A Survey of Image Inpainting Algorithms Based on Transformer Structure

作　　者：柏劲咸 樊瑶[1] 王帅帅 李育博 BAI Jin-xian;FAN Yao;WANG Shuai-shuai;LI Yu-bo(School of Information Engineering,Xizang Minzu University,Xianyang Shaanxi 712000,China)

机构地区：[1]西藏民族大学信息工程学院,陕西咸阳712000

出　　处：《计算机仿真》2024年第8期161-169,174,共10页Computer Simulation

基　　金：国家自然科学基金(62062061)。

摘　　要：近年来,基于Transformer结构的图像修复算法在图像全局结构理解、通用数据集的泛化能力等方面表现出色。然而目前相关研究综述较少,为了进一步推动图像修复问题研究,对Transformer类的图像修复方法进行归纳和分析。首先介绍了Transformer基本原理和框架,其次依据Transformer的结构对采用Transformer的图像修复模型进行分类,分析描述了各方法的改进之处、适用范围和优缺点等,针对不同掩码以及掩码比率,在多种公共数据集上对不同算法进行定量分析和修复效果展示,同时对各种方法的输出多样性进行性能分析。最后总结了相关研究所面临的挑战,并对未来的发展前景和研究方向提出了展望。In recent years,image inpainting algorithms based on Transformer structure have performed well in terms of image global structure understanding,diversified restoration,generalization ability of common datasets,etc.,and the restoration results are more reasonable and diversified.However,there are few reviews of related studies.In order to further promote the study of image inpainting problems,image inpainting methods of Transformer class are summarized and analyzed.Firstly,we introduced the basic principle and framework of Transformer.Secondly,we classified the image inpainting models using Transformer based on the structure of Transformer,and analyzed and described the improvements,applicability,advantages and disadvantages of each method,etc.In addition,quantitative analysis and inpainting effects of different algorithms were demonstrated on various public datasets for different masks and mask ratios,and performance analysis of the output diversity of various methods was also performed.Finally,the challenges faced by the related research are summarized,and the future prospects and research directions are proposed.

关 键 词：图像修复 深度学习 自注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]