Image inpainting based on deep learning - A review图像修复 2021 最新综述

Posted 2022-09-04 墨理学AI

Image inpainting based on deep learning - A review

---

基本信息

• Image inpainting based on deep learning: A review
• 基于深度学习的图像修复：综述
• 本博文乃该论文精简摘录

---

摘要

  图像修复(Image Inpainting)旨在恢复残缺图像中损坏部分的像素特征，在许多计算机视觉应用领域中发挥关键作用。基于深度学习的图像修复技术是当前的一大研究热点，为了深入了解基于深度学习的图像修复相关方法技术，本文对该领域最新研究现状进行脉络梳理和综述。首先对基于深度学习的不同类型神经网络结构修复方法进行归纳总结，然后对其中重要技术改进机制进行分析研究。本文从模型网络结构类型、修复方式等方面对各种算法进行综合评述，选取部分代表性的图像修复方法进行对比和分析。最后对图像修复现阶段问题进行了总结，并对未来的发展趋势和研究方向进行了展望。

---

1. Introduction

Eight common image inpainting applications

---

2. Related works

常规 相关工作总结

2.1. Image inpainting tasks

2.2. Traditional image inpainting

2.3. Image inpainting based on generative network

---

3. Image inpainting methods based on deep learning

从网络结构的视角分三大类

• 3.1. Single-stage inpainting

 3.1.1. Single result inpainting

 3.1.2. Pluralistic inpainting approaches

• 3.2. Progressive image inpainting

 3.2.1. Low-resolution image inpainting

 3.2.2. High-resolution image inpainting

• 3.3. Inpainting based on prior knowledge

 3.3.1. Contour edge guided image inpainting

 3.3.2. Generative prior guided image inpainting

---

4. Image inpainting datasets

NVIDIA Irregular Mask Dataset

六大当前经典图像修复数据集

---

5. Discussion and analysis

Comparing and summarizing the above various types of representative image inpainting methods, we can find：

该部分主要是对第三部分图像修复方法的总结分析

---

6. Conclusions

At present, image inpainting technology has become an important branch in field of vision research. Deep learning image inpainting based on generation network gradually become mainstream method. Researchers have continuously innovated and made great progress in generation model selection, network structure design, introduction of prior guidance, discriminator optimization, loss function optimization, etc.

However, the following problems still need to be solved urgently:

• 4 视频流图像修复，目前基于深度学习的图像修复方法得益于卷积神经网络良好的空间特征提取能力，且大多采用深度卷积神经网络构建网络层。循环神经网络能够在时间序列特征级挖掘数据中的语义信息，在语音和自然语言处理领域有很好的应用。如何有效地结合这两种神经网络来处理视流图像修复将是一个非常有意义的研究方向。

---

致谢

---