Colorization detector

A heuristic detector for colorized black-and-white photos — decides whether a colour image is a true colour photograph (film or digital) or a colorization layered on top of an originally monochrome source. Useful for historians, journalists, museum curators: “is this 1920s-looking colour photograph a real find or Photoshop?”. Three weighted heuristics. (1) Chromatic poverty in Lab space: measures the spread of chromaticity across a* and b*, plus the mean chroma. Real colour photographs are chromatically rich (mean chroma ≥ 15); hand or AI colorization tends to be muted (5-12). (2) “Island” segmentation: measures local a*-b* variance inside 16×16 blocks. Colorization typically shows sharp boundaries between colour regions with near-uniform tone inside the block; a real photo has smooth local gradients. (3) R↔B noise correlation: differences between adjacent pixels in the R and B channels are compared. Real RGB sensors produce independent noise (correlation 0.1-0.3); colorization on top of a B&W master makes the noise identical across channels (correlation > 0.6) — this is the strongest signal. Final score 0-100: below 30 — original colour, 30-60 — uncertain, above 60 — signs of colorization. Raw metrics (σ, mean chroma, blockVariance, noiseCorr) are exposed too for hands-on interpretation. Heuristic, not a verdict: modern AI colorizers (DeOldify, Palette.fm) are specifically trained to produce rich chroma and can land in “uncertain”. Processing happens in the browser, the file never leaves your device.