Colorization as Pretext

Colorization as Pretext is a self-supervised learning task where the model is trained to predict the color channels (a, b in Lab color space) of an image given only the luminance channel (L) — requiring the network to learn semantic understanding to assign plausible colors.

How Does Colorization Work?

• Input: Grayscale image (L channel).
• Output: Predicted a, b chrominance channels.
• Loss: L2 or classification (quantized color bins) loss on predicted colors.
• Paper: Zhang et al., "Colorful Image Colorization" (2016).

Why It Matters

• Semantic Learning: Assigning correct colors requires understanding what objects are — sky is blue, grass is green, skin is flesh-toned.
• Ambiguity: Many objects can be multiple colors (car, shirt) — the model must learn object priors.
• Limitations: The representations are biased toward color-relevant features and may miss texture/shape information.

Colorization is painting by understanding — a pretext task that forces the network to recognize objects and scenes to predict plausible colors.