It becomes like this when you colorize black and white animation with deep learning



Machine learning by multilayered neural network, so-called "Deep learningUsing the power ofgrayscaleThere is an attempt to convert the image of the color to color, and in fact, a movie about what happens when coloring the animation of the black and white era is done on YouTube.

Colorful Image Colorization
http://richzhang.github.io/colorization/


GitHub - pavelgonchar / colornet: Neural Network to colorize grayscale images
https://github.com/pavelgonchar/colornet


Richard Zhang and others are advancing the "deep learning automatic coloring" project to colorize grayscale images with deep learning. Of course, it has not reached the point that it reproduces the original color perfectly, but what is pretty bad is done as an atmosphere.

An example is like this.


Eiji K is also trying this project with animation. For example, the opening of "Cyborg 009" of the first TV anime version broadcasted from April 1968 to September is like this.

Cyborg 009 OP automatic coloring test by OP deep learning - YouTube


The left is the original monochrome version, right is the deep learning automatic coloring version. Cyborg 's clothes are red base and rock is brown, it is painted like it.


003 Francoise. There is unevenness in the painting of face and hair color for a while.


Shimamura Joe. The color of the muffler is properly red.


I feel that the body color of the Shinkansen is a bit more white, but it is painted as gray instead of white in the original version, so it may be correct for deep learning.


Next, the opening of 'Osomatsu-kun' (Phase 1) broadcasted from February 1966 to March 1967. Most recently, "Osomatsu-san" who made a big hit is Kore, so the color is roughly the same image ... ...

Osomatsu-kun OP1 (black & white) Automatic coloring test by deep learning - YouTube


It is like this when 6 people are in stock. The background color is not stabilized. At first, I thought of learning "Osomatsu-san" and painting it like "Osomatsu = red" "Karamatsu = blue", but it was not so indeed.


The uniforms of the six people are painted blue, as imaged.


However, depending on places it will be reddish.


As I think about it because Iyami's clothes have a purple color image.


"Osomatsu-kun" has a second generation opening as well.

Osomatsu-kun OP 2 (black & white) Automatic coloring test by deep learning - YouTube


If the colors clearly come out a bit more ......


Daughan's skin and teeth were painted separately.


And at the end is "Sorcerer Sally" (Phase 1). This work was broadcasted in all 109 stories in December 1966 to December 1968, but in 1967 March broadcasting 17 episodes up to monochrome, from 18 episodes broadcast in April 1967 from color to color Since it is switched, it is possible to check the difference between the deep learning automatic coloring and what color was imaged.

Sally the Witch 1966 Auto Coloring Test by Deep Learning (Colorization) - YouTube


First of all, the appearance of the house. In the color version the roof is red, but with deep learning it is not possible to reproduce that.


Sally sitting on the sofa. In color, the sofa is green, but when it is deep learning it turns to a bluish color. By the way, from the difference of the pattern of the sofa and the picture on the upper right, you can see that you are redrawing a new picture when you make it a color version.


Sally's up. Even with deep learning, clothes are managed red, but the color of the skin does not appear well.


In a drawing by Sally who is flying in the sky, deep learning automatic coloring paints Sally 's clothes clearly brightly, making it easier to see than color version.


In the scenes where the turnips come out, when the deep learning is done, the color of the turnip hair has become reddish. However, the color of the umbrella is delicately painted differently and I feel that I am doing a lot of work.


As you learn more, it may become possible to achieve automatic coloring as per the image.

in Software,   Video,   Anime, Posted by logc_nt