级联式生成对抗网络的全景图像修复  

Cascaded generative adversarial network for panoramicimage inpainting

作　　者：徐嘉悦 赵建平[1] 李冠男 韩成[1] 李华[1] 徐超 XU Jiayue;ZHAO Jianping;LI Guannan;HAN Cheng;LI Hua;XU Chao(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学与技术学院,长春130022

出　　处：《重庆理工大学学报（自然科学）》2024年第8期154-163,共10页Journal of Chongqing University of Technology：Natural Science

基　　金：吉林省自然科学基金项目(20220101134JC)。

摘　　要：为了解决全景图像视场宽、畸变显著等问题,提出了一种级联式生成对抗网络的全景图像修复算法。第一阶段提出了一种双判别器生成对抗网络,通过对等矩形格式的全景图像进行立方体投影转换,对立方体六面图像进行修复,引入PatchGAN作为全局判别器捕获细节信息,局部判别器网络可以保证局部修复结果与周围区域的一致性。第二阶段提出了一种失真感知生成对抗网络,通过矩形混合卷积缓解全景图像失真,判别器引入谱归一化,与第一阶段进行级联以缓解立方体图像边界不连续问题,设计联合损失函数以优化网络修复效果。实验结果表明,所提算法无论从主观视觉评价或是从客观评价指标上均取得了优秀的效果,实现全景图像的有效修复。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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]