Colorizing genuine archival photographs: global palette + targeted scribbles
CNN colorization networks are trained on colour photos converted to grayscale. But real archival photographs (silver halide) don't share the same spectral sensitivities, so the network fails on flags, monuments, well-known objects. I improve the output by combining two types of hints: an automated palette + manual scribbles where it matters.
When the dataset betrays the network
Spectral-sensitivity mismatch: training datasets (ImageNet, Places…) are digital RGB photos converted to grayscale via a weighted average of channels. Genuine silver-halide B&W photos react differently to light: silver halide ≠ Y'CbCr luminance. Grey levels do not map to the same colours.
Practical consequence: for a real archival photo, the predicted colours can be plausible but wrong. A tricolour flag that ends up painted in one flat colour. A monument in the wrong stone, or a military uniform that looks like it belongs to the wrong army.
Building on MCNA 2020: my previous paper proposed a global palette as a hint. Good intuition, but not enough alone for edge cases (unique, identifiable, error-sensitive objects).
Two hints, one colorizer
Global palette + targeted scribbles
The idea: combine two complementary types of hints. The global palette covers most of the frame. Manual scribbles are only used on residual objects where the palette is not enough.
- Automated global palette (inherited from MCNA 2020): a salient palette computed per scene. Good for movie colorization, automatable, precise enough for most scenes.
- Targeted manual scribbles: for regions where the global palette is not enough (flags, monuments, unique objects), the user drops a few localised colour patches. A few pixels are enough.
- No architecture change: hints (palette + scribbles) are injected as additional inputs to the pretrained network. Plug-in on existing colorizers.
- Usage strategy: 80% of colorization is handled by the palette, scribbles are only needed on about 15% of the surface. The ratio shifts with scene complexity.
What this method enables
Visibly better quality
On archival photos vs colorizer alone: sky, foliage and people stay correct, and now flags, monuments and uniforms do too (the exact objects that used to fail).
Controlled interaction cost
The palette handles most of the frame automatically; scribbles are only needed on unique objects. Minutes of scribbling per scene, instead of hours of pixel-by-pixel work.
Open-source compatible
Tested on the most-cited open-source colorizers (Zhang/Isola/Efros, Iizuka/Simo-Serra/Ishikawa). No proprietary model required.
Tailored to silver-halide imagery
The pipeline explicitly addresses the domain shift between converted digital photos and genuine silver-halide B&W, a gap the literature usually skips over.
Four concrete contributions
- Clear identification of the domain shift between training datasets (RGB → grayscale) and archival photos (silver halide). A blind spot in the colorization literature, where grayscale is implicitly assumed universal, which is false for archives.
- Hybrid palette + scribbles methodology: a better trade-off between automation (palette) and targeted quality (scribbles), the two work together rather than competing.
- Practical pipeline for restoring historical photos without retraining a dedicated colorizer: usable directly by restorers with an off-the-shelf open-source colorizer.
- Culmination of the series: EUVIP 2018 (NIR face recognition), EUVIP 2019 (handwriting anomaly detection), MCNA 2020 (colorization palette), IPAS 2020 (palette + scribbles). One thread runs through all four: guiding deep networks with structured hints.
Venue & context
- Published at IEEE IPAS 2020 (International Conference on Image Processing, Applications and Systems).
- Work carried out during my research master's at MIA Laboratory · University of La Rochelle.
- Keywords: CNN, colorization, archival photography, interactive colorization, domain shift.