Work in a perceptual colour space
Explore L*a*b* colour space and predict colour information from the luminance channel.
Computer vision · 2021
Teaching neural networks to colour grayscale images.
Exploring how convolutional networks and conditional generative models can infer plausible colours from black-and-white photographs.
A grayscale image has no single correct colour.
Turning black-and-white pictures into plausible colour images is an underdetermined task: a single grayscale value may correspond to different real-world colours.
That makes architecture selection and evaluation interesting. Visual plausibility is not always captured by a single numerical metric.
Compare different ways to predict colour.
The study explored image-to-image modelling using convolutional architectures, attention-based U-Net components and adversarial training.
Experiment with model families
Compare CNN-based and conditional generator–discriminator approaches, including attention U-Net and GAN variants.
Use complementary metrics
Examine PSNR, SSIM and LPIPS to compare reconstruction quality and perceptual similarity.
Lessons from image generation.
Colourization is inherently ambiguous
A predicted colour image may look plausible even when it differs from the ground-truth image.
Architecture choices involve trade-offs
Convolutional, attention-based and adversarial approaches have different strengths and training considerations.
Metric choice changes the story
PSNR, SSIM and LPIPS provide distinct perspectives rather than a universally sufficient quality measure.
Evaluation should match the real task.
Not every machine learning problem has a single objectively correct prediction. Colourization is a useful example of why evaluation needs both technical measures and thoughtful interpretation.
The project developed experience working with generative models and comparing their outputs systematically.
Read the full report.
The team report documents the architectures, experiments, qualitative examples and evaluation criteria.