[Time-Travel Rephotography] Colorize old photos using machine learning

2022年7月10日日曜日

Artificial Intelligence English

In this article, I will show you how to colorize old black-and-white photographs using the machine learning method Time-Travel Rephotography.

eyecatch
reference: Time-Travel-Rephotography.github.io

Time-Travel Rephotography

Overview

Many historical people were only ever captured by old, faded, black and white photos, that are distorted due to the limitations of early cameras and the passage of time.

Time-Travel Rephotography simulates traveling back in time with a modern camera to rephotograph famous subjects. Unlike conventional image restoration filters which apply independent operations like denoising, colorization, and superresolution, we leverage the StyleGAN2 framework to project old photos into the space of modern high-resolution photos, achieving all of these effects in a unified framework.

A unique challenge with this approach is retaining the identity and pose of the subject in the original photo, while discarding the many artifacts frequently seen in low-quality antique photos. Our comparisons to current state-of-the-art restoration filters show significant improvements and compelling results for a variety of important historical people.

overview
reference: Time-Travel Rephotography

Please refer to this paper for details.

In this article, we will use the above method to colorize any photo.

Demo(Colaboratory)

Now, let's colorize the photo while actually moving it.
The source code is also described in this article, but you can also get it on GitHub below.
GitHub - Colaboratory demo

You can also open it directly in Google Colaboratory from the following.
Open In Colab

Setup environment

Let's set it up. After opening Colaboratory, please set the following and use GPU.

First, get the source code from Github.

%cd /content

!git clone  --depth 1 --recurse-submodules --shallow-submodules \
  https://github.com/Time-Travel-Rephotography/Time-Travel-Rephotography.github.io.git Time-Travel-Rephotography

# for face align
!git clone https://github.com/adamian98/pulse.git

Then install the library.

%cd /content/Time-Travel-Rephotography

!pip3 install -r requirements.txt
!pip3 install --upgrade gdown

Finally, import the library.

%cd /content/Time-Travel-Rephotography

from pathlib import Path
import os

from PIL import Image
from IPython.display import display

from argparse import Namespace
from projector import (
    ProjectorArguments,
    main,
)

This completes the environment setup.

Trained model setup

Next, we will download the trained model.

%cd /content/Time-Travel-Rephotography
!mkdir -p ./checkpoint/encoder

if not os.path.exists('checkpoint/e4e_ffhq_encode.pt'):
  !gdown https://drive.google.com/uc?id=1YLB-3ZCv6FRAWwCHkTvUou_CAvDcb5Pr -O checkpoint/e4e_ffhq_encode.pt
if not os.path.exists('checkpoint/stylegan2-ffhq-config-f.pt'):
  !gdown https://drive.google.com/uc?id=1aSnTVHGNzQ-Eh5rNDOrOc_4SjQbNMq3B -O checkpoint/stylegan2-ffhq-config-f.pt
if not os.path.exists('checkpoint/vgg_face_dag.pt'):
  !gdown https://drive.google.com/uc?id=12BHlsSQM0D8KyXprIc7WmJRLWBWdtF56 -O checkpoint/vgg_face_dag.pt

if not os.path.exists('checkpoint/checkpoint_b.pt'):
  !gdown https://drive.google.com/uc?id=1fXBCyBKNEZfeiI6LEHTEnUg_DjWHJy5r -O checkpoint/encoder/checkpoint_b.pt
if not os.path.exists('checkpoint/checkpoint_g.pt'):
  !gdown https://drive.google.com/uc?id=1YnQEPf7FyfZxAVQg-CpbvdF4otZeTOZh -O checkpoint/encoder/checkpoint_g.pt
if not os.path.exists('checkpoint/checkpoint_gb.pt'):
  !gdown https://drive.google.com/uc?id=1kelQK3pUdeHwu8uVE7A4WZH_EjHxGE6v -O checkpoint/encoder/checkpoint_gb.pt

%cd /content/Time-Travel-Rephotography/third_party/face_parsing
!mkdir -p res/cp

if not os.path.exists('checkpoint/res/cp/79999_iter.pth'):
  !gdown https://drive.google.com/uc?id=1vwm4BcAKISQgcJLvTUcvesk73UIYdDMF -O res/cp/79999_iter.pth

There are many models, but the download will be completed in a minute or two.

Test image setup

Next, download any image you want to colorize.
This time, use wget to download and use a black and white image under the Google Colaboratory environment.

%cd /content/Time-Travel-Rephotography
!mkdir test_imgs

!wget https://www.allcinema.net/img/6/6476/p_40178_01_01_02.jpg -O ./test_imgs/test.jpg

Align the face part from the image.

%cd /content/pulse/

!python align_face.py \
  -input_dir /content/Time-Travel-Rephotography/test_imgs \
  -output_dir  /content/Time-Travel-Rephotography/test_aligns \
  -output_size 512 \
  -seed 12 \

The input image is as follows.

input image


Colorization

Then, we will carry out colorization.
Better results can be obtained by switching spectral_sensitivity depending on the type of black and white photo.

#@markdown 'b' (blue-sensitive), 'gb' (orthochromatic), 'g' (panchromatic)
spectral_sensitivity = "g" # @param ["b", "gb", "g"]
#@markdown Estimated blur radius of the input photo
gaussian_radius =   0.75 # @param {type:"number"}

args = ProjectorArguments().parse(
    args=[str(input_path)], 
    namespace=Namespace(
        spectral_sensitivity=spectral_sensitivity,
        encoder_ckpt=f"checkpoint/encoder/checkpoint_{spectral_sensitivity}.pt",
        encoder_name=spectral_sensitivity,
        gaussian=gaussian_radius,
        log_visual_freq=1000,
        log_dir="log/",
        results_dir="results/"
))
main(args)

Display the generated image.

def get_concat_h_multi_resize(im_list, resample=Image.BICUBIC):
    min_height = min(im.height for im in im_list)
    im_list_resize = [im.resize((int(im.width * min_height / im.height), min_height),resample=resample)
                      for im in im_list]
    total_width = sum(im.width for im in im_list_resize)
    dst = Image.new('RGB', (total_width, min_height))
    pos_x = 0
    for im in im_list_resize:
        dst.paste(im, (pos_x, 0))
        pos_x += im.width
    return dst

im1 = Image.open(f"/content/Time-Travel-Rephotography/results/test_0_{spectral_sensitivity}.png")
im2 = Image.open('/content/Time-Travel-Rephotography/results/test/skin_mask/face_parsing/input.jpg')
im3 = Image.open(f"/content/Time-Travel-Rephotography/results/test_0-{spectral_sensitivity}-G{gaussian_radius}-init(10,18)-s256-vgg1-vggface0.3-eye0.1-color1.0e+10-cx0.1(relu3_4,relu2_2,relu1_2)-NR5.0e+04-lr0.1_0.01-c32-wp(250,750).png")

concat_img = get_concat_h_multi_resize([im1, im2, im3])

display(concat_img)

The output result is as follows.

result

Summary

In this article, I introduced how to colorize old black-and-white photographs using the machine learning method Time-Travel Rephotography.
I sometimes feel instability due to the use of GAN, but it produces natural color photographs.

References

1.  Paper - Time-Travel Rephotography

2. GitHub - Time-Travel-Rephotography.github.io

AIエンジニア向けフリーランスならここがおすすめです

まずは無料会員登録

プロフィール

自分の写真
製造業に勤務する傍ら、日々AIの技術動向を調査しブログにアウトプットしています。 AIに関するご相談やお仕事のご依頼はブログのお問い合わせフォームか以下のアドレスまでお気軽にお問い合わせください。 bhupb13511@yahoo.co.jp

カテゴリ

このブログを検索

ブログ アーカイブ

TeDokology