基于双转移网络的深度特征重排图像修复算法  

Image inpainting algorithm for depth feature rearrangement based on double shift network

作　　者：李海燕[1] 王伟华 郭磊[1] 周丽萍[2] LI Haiyan;WANG Weihua;GUO Lei;ZHOU Liping(School of Information Science and Engineering,Yunnan University,Kunming 650050,China;Editorial Board of Journal of Yunnan University(Natural Science Edition),Yunnan University,Kunming 650050,China)

机构地区：[1]云南大学信息学院,云南昆明650050 [2]云南大学《云南大学学报(自然科学版)》编辑部,云南昆明650050

出　　处：《华中科技大学学报（自然科学版）》2021年第7期74-78,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：云南省高校重点实验室建设计划资助项目;云南省重大科技专项资助项目(2018ZF017)。

摘　　要：为对图像中大面积缺失区域进行合理的结构修复和精细的纹理填充,提出了一种基于双转移网络的深度特征重排图像修复算法.首先提取缺失图像的特征并还原图像纹理细节;然后根据编码器特征估计缺失内容,提出基于解码器特征的内容损失函数,减少全连接层的解码器特征与缺失区域编码器特征之间的距离,保证合成图像语义的准确性和结构的合理性.在公开数据集上,将提出算法与现有经典算法进行对比,结果表明:在大面积缺失图像修复中,提出算法能得到清晰、细节细腻且视觉合理的结果.To repair the large missing area of images with reasonable structure and pad fine-detailed textures,an image inpainting algorithm based on double shift net by depth feature rearrangement was proposed.First,features of the damaged region were extracted and the texture details were restored.Then,the missing parts were estimated based on encoder features.A content loss function was put forward based on decoder feature to minimize the distance between the decoder feature of the fully connected layer and the ground-truth encoder feature of the missing parts and to ensure the accuracy of the semantics of the synthesized image and rationality of the structure.On the Places datasets,the proposed algorithm was compared with the existing classic algorithms.The experimental results show that the proposed algorithm can obtain clear,fine-detailed and visually reasonable results in the inpainting of large missing areas images.

关 键 词：图像修复 双转移网络 特征重排 解码器特征 内容损失函数 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]