Show me the solution #1

[victorcmoura]

Problem: Colorizing black and white images

• Default solution: Colorize by hand, with photo editors
• Proposed solution: Colorize automatically by previous experiences (based on a tutorial by Emil Wallner)

Techincal explanation

1. Getting the dataset
   • Public image dataset made available by Emil
2. Convert colored images from RGB to Lab
   • L: Lightness (Black and white picture)
   • a: green-red
   • b: blue-yellow

This way, instead of predicting 3 channels (red, green, blue), there remains only two.

3. Max-pooling
   • Technique used to increase detail density.

4. Fusion architecture

   • Encoder (Smaller image patches)
   • Decoder (Global priors)
   • Fusion

5. Software Engineering improvements

   • Instead of a jupyter notebook, the code has been transformed into a Python package
   • Pipenv has been used as a configuration manager
   • Trained models have been saved in .h5 files, to save time
   • A CLI has been implemented to enhance the usability

Results

Test training (i7 8750H, 6 cores, 16GB RAM)

Dataset size: 999 images
Epochs: 10
Step: 2 (500 images per batch)

Time elapsed: ~40m

Local training (i7 4500U, 4 cores, 16GB RAM)

Dataset size: 20 images
Epochs: 50
Step: 2 (10 images per batch)

Time elapsed: ~10h

Google Cloud 8 cores, 30GB RAM

Dataset size: 5000 images
Epochs: 1000
Step: 20

Time elapsed: 50h

Google Cloud 8 cores, 30GB RAM

Dataset size: 400 images
Epochs: 1000
Step: 20

Time elapsed: 48h12min

Lessons learned

• Deep learning is $$$$$
• General networks are harder to achieve
• Dependencies errors or mismatch
• Difficulty to use GPU to train the model

Future Works

• Find better parameters from training
• Use GPU for Training
• Use GAN (Generative Adversarial Network) for better results
• Create CI/CD for releases