Bachelor thesis

• B.Sc
  • data/
    • Dataset examples
  • evaluation/
    • Colorized examples
  • implementation/
    • Colorization.ipynb as Jupyter Notebook
  • model/
    • Model summaries

Title

Automatic colorization of gray-value images based on generative algorithms of machine learning and deep neural networks

Abstract

Generative models create - in contrast to object recognizing models – new con-tent instead of just differentiating between familiar instances. This reveals novel possibilities of data synthesis, which are required whenever it is not useful or even impossible to predict unknown content. The automatic colorization of gray-value images is a prime example for the use of generative models, like a GAN (Generative Adversarial Network), which assigns as part of this work plausible color information to colorless patterns. Deep convolutional neural networks are used, which handle the colorization as a pixel-to-pixel transformation through machine learning. As a result, historical black and white images can be enhanced by applying a suitable color restoration.

Baseline

Based on Pix2Pix by Isola et al. from Image-to-Image Translation with Conditional Adversarial Networks

Data

Trained on Places365 by Zhou et al. from Places: A 10 million Image Database for Scene Recognition

Model

• Generative Adversarial Network
  • Architecture
    • U-Net as Generator
    • PatchGAN as Discriminator
  • Loss
    • Binary crossentropy
    • Least absolute deviations, L1
  • Optimiser
    • Adam

Implementation

Release	Date
1.0.0	31.01.20

Software

Software	Version
CoLab	1.0.0
Keras	2.2.4
Python	3.6.9
TensorBoard	2.1.0
TensorFlow	2.1.0
TensorFlow Addons	0.6.0

Acknowledgement

Thesis supported by Faculty of Information, Media and Electrical Engineering at TH Köln - University of Applied Science

Implementation relies heavily on TensorFlow and Keras

Training process borrows from TensorFlow Tutorial

Hyperparameter search based on TensorBoard Guide