Final Project for the CAS in Advanced Machine Learning (UniBern 2023-2024)
Authors:
Lorenz Joss and Ruben Lopez
Image colorization techniques aim to transform a grayscale image by assigning colors, so it becomes visually similar to our perceived reality. This task is challenging since a number of features need to be taken into account: global color richness balance, visual harmony, the conformity to object-level semantics, etc. while minimizing the visual artifacts such as color leakage or incomplete colorization. Current advances in machine learning (ML) and artificial intelligence (AI) have provided a huge contribution to automate this task. Historical aerial images is one of the many expertise domains that can profit from deep-learning base colorization applications. In this project we aim to use an existing model based on U-Net architecture and retrain it with historical aerial images from Switzerland.
Example of image colorization results using diferent models.
This repositry contain the script we used to perform the following task
- data retrieval
- data preparation
- Model training usign containers to the HPC UBELIX
- Model metrics and history plots
- Model inference with color or grey images
The ready-to-feed-model data and model's weights from this project can be found in the following links:
If you wish to look at details on how the data was preprocessd before preparing and feed it to the models, please see the script for data preparation.
Please report issues or comments via GitHub Issues.