文献类型：期刊文献

中文题名：基于自适应跨域距离一致的有限数据图像修复

英文题名：Image inpainting under limited data based on adaptive cross-domain distance consistency

作者：厉嘉琦[1];肖婷[1];杨孟平[1];王喆[1,2]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

年份：2024

卷号：45

期号：12

起止页码：3674

中文期刊名：计算机工程与设计

外文期刊名：Computer Engineering and Design

收录：CSTPCD;;北大核心:【北大核心2023】；

基金：上海市“科技创新行动计划”基金项目(20511100600、21511100800);国家自然科学基金项目(62076094)。

语种：中文

中文关键词：图像修复;深度学习;有限数据;预训练;域自适应;公共表征;相对距离

外文关键词：image inpainting;deep learning;limited data;pre-training;domain adaptation;public representation;relative distance

摘要：针对现有深度学习图像修复方法在有限数据场景下存在的修复质量差、多样性弱等问题,在预训练基础上提出一种域自适应的方法,迁移源域中可学习的知识,补充训练中所需的信息。对源域和目标域中的特征信息进行探究,发现特征中的结构信息可作为公共表征,为域自适应提供学习的来源;为更有针对性地完成域自适应过程,提出一种自适应跨域距离一致性损失,自适应地调节损失权重,保留更多与目标域接近的源域相对距离,完成对源域知识的学习。实验结果表明,所提方法能有效提升修复质量和真实性,并且具有良好的泛化性。
To address the issues of poor results and limited diversity in existing deep learning-based image inpainting methods,a domain adaptation approach was proposed based on the pre-training.The learnable knowledge was transferred from the source domain to supplement the required information during training.Structural information in the feature space was explored as the public representation between the source and target domains.An adaptive cross-domain distance consistency loss was proposed to preserve relative distances between the source and target domains by adaptively adjusting the loss weight.Experimental results demonstrate that the proposed method effectively improves the inpainting quality and realism,and exhibits good generalization performance.