deepfake 資源總結

 

1. https://zhuanlan.zhihu.com/p/34042498   深度解密換臉應用Deepfake

2. 在 1 里面提到的PixelShuffle，具體見參考3：

https://mathematica.stackexchange.com/questions/181587/how-to-define-a-pixelshuffle-layer

一邊Upsample一邊Convolve：Efficient Sub-pixel-convolutional-layers詳解

https://oldpan.me/archives/upsample-convolve-efficient-sub-pixel-convolutional-layers

正常情況下，卷積操作會使feature map的高和寬變小。但當我們的stride=(1/r) < 1時，可以讓卷積后的feature map的高和寬變大——即分辨率增大，這個新的操作叫做sub-pixel convolution，具體原理可以看PixelShuffle《Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network
》的論文。

pixelshuffle算法的實現流程如上圖，其實現的功能是：將一個H×W的低分辨率輸入圖像（LowResolution），通過Sub-pixel操作將其變為rH x rW的高分辨率圖像（High Resolution）。

在1中提到的PG-GAN

3. http://openaccess.thecvf.com/content_cvpr_2017/papers/Ledig_Photo-Realistic_Single_Image_CVPR_2017_paper.pdf

4. PyTorch學習筆記(10)——上采樣和PixelShuffle

https://blog.csdn.net/g11d111/article/details/82855946

5. faceswap blog

https://blog.csdn.net/weixin_41965898/article/details/84930788

參考：

1. CNN概念之上采樣，反卷積，Unpooling概念解釋 

https://blog.csdn.net/g11d111/article/details/82350563

2. Visualizing and Understanding Convolutional Networks

https://arxiv.org/pdf/1311.2901v3.pdf

3. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network

【這篇文章的核心—Efficient Sub-pixel Convolution】

https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Shi_Real-Time_Single_Image_CVPR_2016_paper.pdf

4.Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

http://openaccess.thecvf.com/content_cvpr_2017/papers/Ledig_Photo-Realistic_Single_Image_CVPR_2017_paper.pdf

5【超分辨率】Efficient Sub-Pixel Convolutional Neural Network【Sub-Pixel / PS: periodic shuffling】

https://blog.csdn.net/shwan_ma/article/details/78440394

6. PixelShuffle的含義