基于并行对抗与多条件融合的生成式高分辨率图像修复

Generative High-Resolution Image Inpainting with Parallel Adversarial Network and Multi-condition Fusion

现有的图像修复算法经常会有伪影、语义不准等问题出现,对于缺失较大、分辨率较高的图像,修复效果有限.为此,文中提出基于并行对抗与多条件融合的二阶图像修复网络.首先,利用改进的深度残差网络对缺失图像进行生成式像素填充,并利用第一阶对抗网络补全边缘.然后,提取填充图颜色特征,融合补全边缘图,将融合图作为第二阶对抗网络的条件标签.最后,通过带上下文注意力模块的第二阶网络得到修复结果.在多个数据集上的实验表明,文中算法可获得较逼真的修复效果.

Regions with artifacts and semantic inaccuracy are often caused by existing image inpainting algorithms.Moreover,the inpainting effect is limited for images with large missing regions and high-resolution.Therefore,a two-stage image inpainting approach based on parallel adversarial network and multi-condition fusion is proposed in this paper.Firstly,an improved deep residual network is utilized to fill the corrupted image.The first-stage adversarial network is employed to complete the image edge map.Next,the color feature of the filled image is extracted and fused with the completed edge image.Then,the fused image is applied as the condition label of the second-stage adversarial network.Finally,the inpainting result is obtained by the second-stage network with a contextual attention module.Experiments on multiple public datasets demonstrate that realistic inpainting results can be obtained by the proposed approach.