AI 修復(fù)是如何給黑白影像上色的？

How does AI repair color black and white images?

 

Technical Principle

AI 上色的原理是什么？那我們就需要介紹一種深度學(xué)習(xí)網(wǎng)絡(luò)架構(gòu)了，它就是 GAN（這里不是粗話）。GAN 不是干飯人的干，而是生成對抗網(wǎng)絡(luò)（英語：Generative Adversarial Network，簡稱 GAN）。當(dāng)然，太復(fù)雜的技術(shù)講解可能會讓讀者迷惑，于是我找到一張很直白的原理圖。

What is the principle of AI coloring? Then we need to introduce a deep learning network architecture, which is GAN (not foul language here). GAN is not the job of a fan, but a generative adversarial network (English: Generative Adversarial Network, referred to as GAN). Of course, too complicated technical explanations may confuse readers, so I found a very straightforward schematic.

GAN 網(wǎng)絡(luò)分兩部分，一個是生成器（Generator），一個是鑒別器（Discriminator）。生成器通過對圖像上色，然后交給鑒別器。鑒別器判斷這一個圖片看起來真不真，如果覺得假，鑒別器會返回「修改意見」，讓生成器重新試試，直到鑒別器覺得足夠真了。如果你覺得還不好懂，我再打個比方，這就好像美術(shù)老師指導(dǎo)學(xué)生畫畫的過程，一開始學(xué)生畫出來的不夠好，老師指出，學(xué)生嘗試改改，老師再檢查，再給意見，直到老師滿意。

The GAN network is divided into two parts, one is the generator and the other is the discriminator. The generator colors the image and then passes it to the discriminator. The discriminator judges that this picture looks true or not. If it feels false, the discriminator will return the "modification opinion" and let the generator try again until the discriminator feels it is true enough. If you think it’s not easy to understand, let me make an analogy. It’s like the process of an art teacher instructing students to draw. At the beginning, the students did not draw well enough. The teacher pointed out that the students tried to make changes, the teacher would check again, and then give comments. Until the teacher is satisfied.

這就是一張圖的上色過程。而視頻是一幀幀畫面組成的，給視頻上色可以理解為通過這個網(wǎng)絡(luò)架構(gòu)給視頻里的每一幀上色。不過沒有這么簡單，畢竟視頻一秒鐘幾十幀，一幀幀上色有點(diǎn)慢，而且每一幀之前可能會出現(xiàn)上色效果不一致。所以有的架構(gòu)會針對細(xì)節(jié)調(diào)整。例如 DeOldify 采用了 NoGAN（一種新型 GAN 訓(xùn)練模型），用來解決之前 DeOldify 模型中出現(xiàn)的一些關(guān)鍵問題。例如如視頻中閃爍的物體

This is the coloring process of a picture. The video is composed of frames, and coloring the video can be understood as coloring each frame in the video through this network architecture. But it is not that simple. After all, the video has dozens of frames per second, and the coloring of one frame is a bit slow, and the coloring effect may be inconsistent before each frame. So some architectures will be adjusted for details. For example, DeOldify uses NoGAN (a new type of GAN training model) to solve some of the key problems in the previous DeOldify model. E.g. flashing objects in the video

一般來說，給人上色會更接近實(shí)際情況些。因?yàn)槿说哪w色比較有限，判別器里已經(jīng)學(xué)習(xí)過人臉的顏色可能是哪些，轉(zhuǎn)換成灰度圖像后對應(yīng)什么顏色值，所以 AI 不太可能會給黑白的人像涂成綠色臉。

Generally speaking, coloring people will be closer to the actual situation. Because human skin color is relatively limited, the discriminator has already learned what the color of the face may be, and what color value it corresponds to after converting it into a grayscale image, so it is unlikely that AI will paint a black and white portrait with a green face.

 

From website: https://daily.zhihu.com/