Image inpainting based on multi-scale stable-field GAN
Xueyi Ye, Maosheng Zeng, 伟杰 孙, 凌宇 王, 知劲 赵 · Scientia Sinica Informationis · 2022
Generative adversarial networks (GANs) have shown their potential to inpaint large missing regions and generate plausible semantic results in image-inpainting tasks. However, GAN-based methods often ignore the semantic consistency and feature continuity of missing regions, and lack the perceptibility of features at multiple scales. To address these issues, we propose an image-inpainting model based on GAN with a multi-scale stable field. Motivated by the U-Net architecture, its generator embeds the stable field operator into skip connections to fill the missing region in the feature map of the encoder and thus maintain the semantic consistency and feature continuity of the missing region. Further, as the benefit is gradually enhanced by multi-scale fusions, the feature information transferred by the skip connections is not obtained from only a single feature map. The model is generally enabled to perceive the semantic information of the high-level features of an image. The experimental results show that the proposed model outperforms other classical image-inpainting methods in the face and natural scene datasets.