计算机视觉——图像修复综述篇

1. Deterministic Image Inpainting 判别器图像修复

1.1. sigle-shot framework

(1) Generators

(2) training objects / Loss Functions

1.2. two-stage framework

2. Stochastic Image Inpainting 随机图像修复

2.1. VAE-based methods

2.2. GAN-based methods

2.3. Flow-based methods

2.4. MLM-based methods

2.5. Diffusion model-based methods

3. text-guided image inpainting ⽂本引导的图像修复

4. Inpainting Mask 掩码机制

(1) regular mask

(2) irregular mask

5. Loss Function 损失函数

6. Dataset 图像修复领域数据集

(1) faces(CelebA & CelebA-HQ)

(2) real-world encountered scenes(Places2)

(3) street scenes(Paris)

(4) texture(DTD)

(5) objects (ImageNet)

7. Evaluation Protocol 评估指标

7.1. pixel-aware metrics

7.2. (human) perception-aware metriics

8. Performance Evaluation 表现评估

8.1 Representative Image Inpainting Methods

8.2 Loss Functions

9. Inpainting-based Application 基于图像修复的领域应⽤

(1) Object Removal

(2) Text Editing

(3) Old Photo Restoration

(4) Image Compression

(5) Text-guided image editing

Reference

1. Deterministic Image Inpainting 判别器图像修复

1.1. sigle-shot framework

(1) Generators

1) mask-aware design

2) attention mechanism

3) multi-scale aggregation

4) transform domain

5) encoder-decoder connection

6) deep prior guidance

(2) training objects / Loss Functions

1) Pixel-wise reconstruction loss

2) perceptual loss

3) style loss

4) adversarial loss

5) prevalent training objectives

1.2. two-stage framework

(1) coarse-to-fiine methods

(2) structure-then-texture methods

2. Stochastic Image Inpainting 随机图像修复

2.1. VAE-based methods

2.2. GAN-based methods

2.3. Flow-based methods

2.4. MLM-based methods

2.5. Diffusion model-based methods

(1) sample stratage design

(2) computational cost reduction

3. text-guided image inpainting ⽂本引导的图像修复

4. Inpainting Mask 掩码机制

(1) regular mask

(2) irregular mask

5. Loss Function 损失函数

同1-1.1-(2) training objects

6. Dataset 图像修复领域数据集

(1) faces(CelebA & CelebA-HQ)

(2) real-world encountered scenes(Places2)

(3) street scenes(Paris)

(4) texture(DTD)

(5) objects (ImageNet)

7. Evaluation Protocol 评估指标

7.1. pixel-aware metrics

focus on the precision of reconstructed pixels

(1) l1 error

(1) l2 error

(3) PSNR(peak signal-to-noise ratio)

(4) SSIM(the structure similarity index)

(5) MS-SSIM(muti-scale SSIM)

7.2. (human) perception-aware metriics

the visual perception quality

(1) FID(Frechet Inception diistance)

(2) LPIPS(learned perceptual image patch similarity)

(3) P/U-IDS(pair-unpair Inception discriminative score)

8. Performance Evaluation 表现评估

8.1 Representative Image Inpainting Methods

(1) Models: RFR, MADF, DSI, CR-Fill, CoModGAN, LGNet, RePaint

(2) Dataset: CeleBA-HQ, Places2

(3) Mask: M1, M2, M3, M4, M5, M6

(4) Metrics: l1, PSNR, SSIM, MS-SSIM, FID, LP-IPS

(5) Loss: pixes reconstruction loss, perceptual loss, resnetpl loss, style loss, stylemeanstd,

percept-style loss, lsgan