High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis

1. Motivation

  • 传统的方法假设图像中不同的patch之间的相似性高;
  • 生成图像的纹理细节不好,且不能处理高分辨率的图像。

2. Approach

2.1 Network Structure

High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第1张图片 Framework Overview.
High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第2张图片 The network architecture for structured content prediction.

网络分为两部分,Conten Network和Texture Network。Conten Network采用编码和解码结构,用于生成高级的语义结构;Texture Network用来生成纹理细节。

2.2 Loss function

  • The Joint Loss Function:

High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第3张图片

High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第4张图片

  • The Content Network:

High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第5张图片

3. Discussion

这篇文章中Conten Network可以在语义和整体结构上提供很强的先验知识,Texture Network也可以很好的生成一些细节。但是这篇文章中的方法不能处理复杂的情景,下图展示的是一些失败的例子,此外运行的时间较长。

High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis_第6张图片

4. References

【1】Yang, Chao, et al. "High-resolution image inpainting using multi-scale neural patch synthesis." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.

你可能感兴趣的:(深度学习,计算机视觉)