重庆理工大学学报2024，Vol.38Issue(15)：154-163,10.DOI:10.3969/j.issn.1674-8425(z).2024.08.018
级联式生成对抗网络的全景图像修复
Cascaded generative adversarial network for panoramic image inpainting
摘要
Abstract
To address problems as the wide field of view and significant distortion in panoramic images,a panoramic image inpainting algorithm using a cascaded generative adversarial network is proposed in this paper.First,a dual-discriminator GAN is introduced to perform cubic projection transformations on equi-rectangular format panoramic images and restore the six faces of the cubic image.It incorporates a PatchGAN as the global discriminator to capture fine details,while a local discriminator network ensures the consistency of local restoration results with the surrounding areas.Then,a distortion-aware Generative Adversarial Network is introduced to mitigate distortion in panoramic images by using rectangular mixed convolutions.The discriminator employs spectral normalization and is cascaded with the first stage to mitigate issues of discontinuities at cubic image boundaries.A joint loss function is designed to optimize the network's restoration performance.Our experimental results demonstrate the proposed algorithm dilivers excellent performances in both subjective visual evaluation and objective assessment metrics,effectively restoring panoramic images.关键词
全景图像/图像修复/生成对抗网络/双判别器/投影转换/混合卷积Key words
panoramic images/image inpainting/generative adversarial network/dual-discriminator/projection transformation/hybrid convolutions分类
计算机与自动化引用本文复制引用
徐嘉悦,赵建平,李冠男,韩成,李华,徐超..级联式生成对抗网络的全景图像修复[J].重庆理工大学学报,2024,38(15):154-163,10.基金项目
吉林省自然科学基金项目(20220101134JC) （20220101134JC）
