Remove the mosaic, there's a way, attached with the running tutorial

Eliminate mosaic and turn into high-definition portrait in seconds , Make the blurred picture clear in seconds .

ai Technology is becoming more and more powerful .

But now ai technology , It can really eliminate mosaic completely , 100% restore photos ？

Actually , Mosaic elimination algorithm PULSE, stay 2 It was released years ago .

Make up the coding place by algorithm , Help us restore photos .

Even the pores on the face 、 Hair can be restored .

Up to now , The technology of eliminating mosaic has matured , Not two weeks ago , Open source a new algorithm MAE.

This is more powerful , Even if the screen is blocked for more than 90%,ai Will also find ways to help us recover .

I tried it out , The following picture is used .

After occlusion , We can roughly see what this is , But I don't know what the eyes look like .

Run it for a while , The effect simulated by the algorithm is as follows ：

And the original picture is like this ：

It is basically restored .

Let's take a look at the principles of these algorithms .

Mosaic is actually a low pixel image .

A clear picture , It's just the eye area , There may be 100 Pixel , And turn it into mosaic , This area may only be 3 Pixel .

So the picture will be blurred , I can't see what the original picture is .

I want to restore the photos , It is necessary to base on this 3 Pixel , The brain makes up some features , Make up again as 100 Pixel .

This is what the algorithm does , Automatic brain filling some nonexistent features , Like wrinkles 、 Hair, etc .

Can these algorithms really help us 100% restore photos ？

I took my own ID photo and tried it .

The first is the original , The last one is the effect of the final restoration .

that ——

Although they also have nose and eyes .

But compared with the original photo , It's irrelevant .

The photo generated at last is only made up by the algorithm , Maybe it's just a face that looks real .

But I want to reconstruct and restore the portrait through mosaic , So far , It's impossible .

But restore some things that are not obvious in their individual characteristics , Like furniture 、 Animal, etc. , The accuracy is relatively high .

For example, this little fox ,MAE The algorithm restoration is very successful .

So how to run these algorithms ？

We use the first algorithm PULSE give an example .

First step ： Download code

Project address ：

https://github.com/adamian98/pulse

The second step ： Set up the environment

Under the root directory of the code file , There is one pulse.yml file , The environment configuration information is all here .

We can directly run the following code to create a virtual environment , And according to pulse.yml File configuration environment .

conda env create -n pulse -f pulse.yml

But I install it myself , The following error was reported ：

Solving environment: failed 
ResolvePackageNotFound:***

This is because , The environment configuration file is exported from another computer , Not suitable for our computer configuration , Deleting the specific information of the package can solve the problem , That is, delete the contents after the second equal sign of the package name .

such as ：

- blas=1.0=mkl
- ca-certificates=2020.1.1=0

Change to ：

- blas=1.0
- ca-certificates=2020.1.1

also dlib The installation of the library also encountered problems , You need to install cmake, Install again dlib.

pip install cmake
pip install dlib

If the installation is still unsuccessful , You can put dlib19.19.0 Download the version , Install... Locally .

I will dlib19.19.0 Version file , Uploaded to my baidu disk .

Download address （ Extraction code ：6666）：

https://pan.baidu.com/s/16KHEdZ0KD_pQPRRiGuq5Ew

Finally, install the local file .

pip install dlib-19.19.0-cp38-cp38-win_amd64.whl.whl

The third step ： Run the model

The project provides us with a pre training model , Need to climb over the wall to download .
I have already downloaded it , Uploaded to my baidu disk .

Download address （ Extraction code ：6666）：

https://pan.baidu.com/s/16KHEdZ0KD_pQPRRiGuq5Ew

Under the root directory of the code file , Create two folders , Named as ：cache、realpics .

Download the above , Pre training model （ Three files ）, Put it in cache In the folder .

Then put a portrait photo in realpics In the folder , Let's take the following picture as an example ：

Run the following statement first , Reduce the resolution of the picture .

python align_face.py

The generated image will be placed in input In the folder .

Last run run.py file

python run.py

Will be in runs Create brain mapping under folder .

They are all blonde , But not much like ——

If this error is encountered during operation ：

Could not find a face that downscales correctly within epsilon

There are two ways to solve ：
1. Increase iterations

python run.py -steps=5000

2、 increase eps

This is a problem , Because L2 The loss is greater than eps, We increase by eps Value , You can avoid this mistake

stay run.py Of documents 39 That's ok .
Source code ：

parser.add_argument('-eps', type=float, default=2e-3, help='Target for downscaling loss (L2)')

Modified into ：

parser.add_argument('-eps', type=float, default=8e-2, help='Target for downscaling loss (L2)')

In this way, there is basically no problem .

I have written down all the problems encountered during the project , I hope you can run smoothly ~

If you want to run the new MAE Algorithm .

The project address is ：

https://github.com/facebookresearch/mae

The project provides Colab, To log in Google Account can be run , If you can log in , You can directly experience the algorithm effect online ：

https://colab.research.google.com/github/facebookresearch/mae/blob/main/demo/mae_visualize.ipynb

wx Search for 【 Meow Ningyi 】 Read the article for the first time ~