Using Neural Networks to De-censor Anime Girls

A Guide to Implementing a Convolutional Neural Network for image prediction

Oct 20, 2022

Note: This AI is literally just a linear regression, but on stereoids.

For this post, we’ll be using a library called DeepCreamPy (DCP), to de-censor some anime girls. After reading this post, you too will learn how to become a man of culture. @Deeppomf was the original creator of DCP. As of right now there are several forks of it on GitHub. We’ll be using the fork made by gguilt.

Credit goes to @BowTiedCoffee for helping me grasp Photoshop’s Content Filler. Credit also goes to @hal_jpeg, for helping making my code a lot more readable. Also, thanks to @BowTiedNightOwl for letting me use his avatar.

1 - Setting Up DCP

1.1 Installing Dependendencies

DCP runs on Python 3.7. The code for this neural network is super outdated. You need to make sure you grab Python 3.7, and it has to be 64 bit compatible. It has the following dependencies, so go ahead and grab the libraries below. If you don’t know how to add libraries to your code, go here.

absl-py==0.7.1
astor==0.8.0
gast==0.2.2
google-pasta==0.1.7
grpcio==1.22.0
h5py==2.9.0
Keras-Applications==1.0.8
Keras-Preprocessing==1.1.0
Markdown==3.1.1
numpy==1.17.0
opencv-python==4.1.0.25
Pillow==6.2.2
protobuf==3.9.1
PySide2==5.13.0
scipy==1.3.0
shiboken2==5.13.0
six==1.12.0
tensorboard==1.14.0
tensorflow-estimator==1.14.0
tensorflow==1.14.0
termcolor==1.1.0
Werkzeug==0.15.5
wrapt==1.11.2

1.2 Downloading the Model Weights

For the AI to run , it will need to look at several pictures of anime girls (censored & de-censored). Then, it will need to study these pictures to see what’s the common underlying pattern in all them. , we don’t have all day, so we will download the finalized model. Link to the finalized model is here.

1.3 Download & Setup DCP

One thing you’ll notice very is if you tried to look for DCP in Pycharm, it doesn’t exist. What we will have to do is download the repository, and then install it. Instructions Below:

1 - Download the The DeepCreamPy Code as a zip.

2 - Extract the contents of the above zip file to wherever you want your code to run from

3- Now, Download the Model weights file, and extract this to your /models folder.

Congrats, you are now ready to run this algorithm. Just grab a picture, then censor it in pure green (0,255,0)

2 - Running DCP

For this example, I’ll censor @BowTiedNightOwl. His avatar has a lot of curves. You also need to ensure that your picture is in a .png format.

2.1 Censoring it manually

After this, just throw in some censor bars onto your picture.

I am on Windows, so paint can do a good enough job at this.

2.2 Putting it In the right folder

Now go ahead, and save the above picture into the folder called decensor_input.

2.3 Running the Algorithm

Now, run the Python file called decensor.py, and enjoy the output.

3 - Analysis of the Neural Network

3.1 How This Algorithm worked

This algorithm works off of a model called a Convolutional Neural Network (CNN). These neural networks make a simple square from the top left of the image. Then, that square moves row by row, until it reaches the bottom right side of the image.

In this specific case, the neural network was asked to find the green censor bars. Once it did that, then it had to predict what was behind the green censored bar. Yes, this AI is a super complex linear regression.

You can see a simple diagram of this up above. This is how most image detection AIs also work.

3.2 DCP vs Photoshop

Photoshop has an option called Content Filler, and I wanted to compare the two side by side:

@BowTiedNightOwl

The Neural Net was smart enough to figure out how to draw the red nose. Photoshop completely flopped that.

@BowTiedCoffee:

They look almost identical here.

Obligatory Anime Girl

Here, you can see the AI shine. It studied a huge amount of anime girls. Needless to say it would do a better job than photoshop. Big difference is in the eye. Both the AI & Photoshop fumbled the ball there. The AI completely flopped the missing eye, and photoshop did a horrible blend.

Summary:

In the above images, it seems that the neural network is aggressive in it's predictions. When you look at the NightOwl’s red nose, the neural net nailed the missing nose. The neural network is aggressive in that it tries to predict what the lines should look like. But, photoshop tends to try to use the surrounding colors for a soft simple blend.