基于生成对抗网络的分级联合图像补全方法

Hierarchical Joint Image Completion Method Based on Generative Adversarial Network

已有的图像补全工作大都基于规则的、区域较小或者有足够上下文信息的待补全区域。当待补全区域面积较大时,由于上下文信息的缺失及生成对抗网络(GAN)训练的不稳定性,往往会产生模糊或失真的补全结果。尤其是当缺失区域位于图像边缘位置时,补全结果会出现较大的空白及伪彩色。基于以上情况,在已有的基于GAN的补全方法的基础上提出一种分级联合图像补全方法,并针对GAN训练不稳定的问题对网络结构做出了改进。一方面改善了由于缺失区域面积较大产生的补全结果有空白生成的问题,从而使补全结果的纹理细节更加真实、清晰;另一方面使得对抗网络训练更加稳定,抑制了伪彩色的生成。实验结果表明分级联合图像补全方法取得了更好的补全结果。

Existing image completion work is mostly based on missing regions with regular, small area or sufficient context information. When the area of the region to be completed is relatively large, the completion work tends to be blurred or distorted due to the lack of context information and the instability of the generative adversarial network(GAN) training. Especially, if the missing area is located at the edge of the image, the final completion result will have large blank area or pseudo color. To solve the above two problems, the method of hierarchical joint image completion on the basis of GAN is proposed, and the network structure is improved to address the problem of unstable GAN training. On the one hand, it overcomes the problem of the generation of blank area in completion results due to the large missing area, thereby producing more realistic and clear texture details. On the other hand, it makes the adversarial network training more stable and suppresses the generation of pseudo-color. The experimental results demonstrate that the proposed method achieves better completion results.