Neural networks against censorship hentai

From the translator: This funny repository has been in the trends of the githaba for a week now, and now it has an enchanting amount of 5000 stars and 400 forks - so I thought it would be interesting to translate an article about it, although it is short, and the subject matter for Habr is not very standard.
If interested, the original article (link near the title) has links to examples of the program.

In Japan, pornographic images with penises and vaginas are considered illegal. Since the early 1980s, hundreds of people have been arrested for distributing similar materials online.

In order to circumvent this prohibition, artists who paint hentai (pornographic anime and manga genre) usually paint small stripes or blur where there is forbidden content. It turns out that the authors often put a lot of effort into the images of the genitals that will never see the light in Japan (translator's note: yes, that’s what it sounds in the original).

In order to restore images, a machine learning researcher under the name “Deeppomf” wrote an algorithm that uses neural networks to circumvent censorship. He called his program DeepCreamPy.

Anyone can download the code and try this program - it gained a lot of fame on Reddit, and some posted the same images that they got as a result of processing. Over the week, the program scored over 500 downloads.

At the moment, DeepCreamPy can remove censorship in the illustrations with penises and vagin. She's not working on videos, nipples and anus yet. Before launching the program, the user just needs to mark in green the corresponding zone on the image.

Deeppomf wrote to me in an email that the creation of DeepCreamPy is the answer to decades of censorship by the Japanese government. Article 175 of the Penal Code of Japan prohibits the distribution of “obscene” materials (and hentai clearly falls into this category) and requires that pornography be at least partially painted over with black stripes or blurred mosaics.

“DeepCreamPy is designed to counteract increasing censorship,” said Deeppomf. “The Japanese government, especially the Supreme Court of Japan, shows that they are not going to soften their demands. So change must come from the outside. ”

He began to think about the project in the summer of 2016, but said that he was not the first to use neural networks to remove censorship.

“But I was the first to be motivated enough to implement this idea, skillful enough to collect data and train the neural network, and not career-oriented enough to work on something that would fit my resume better.” - he added as well.

In just a few months, Deeppomf collected a sample of more than 100,000 uncensored, uncensored images to train the algorithm. The sample includes yaoi, yuri, feet, and “other fetishes,” as he said. Gouraud and faecal pornography were not included in the sample, as they could lead to undesirable results.

“I did not collect my own porn collection, so my personal attachments don't matter. I have not seen more than 95 percent of the images, because it would take too long to study each, ”commented Deeppomf.

So far, the algorithm does not always produce correct results - at least in those examples provided by users - black stripes are removed quite easily, but there are problems with mosaics. However, Deeppomf continues to improve the algorithm, and also wants to add a new interface and the ability to work with black and white images.

“Artificial intelligence has just begun to penetrate the porn industry,” says Deeppomf. "It's only the beginning".

Links

• Project repository
• Thread on reddit
• The image of the mermaid in the example is taken from the Shurajo & AVALANCHE Game Studio, the original is published under the license CC BY 3.0 License and is located here .