Converts color images to grayscale while preserving the contrast between colors that ordinary luminance conversions flatten out.
use decolorize::{decolorize, decolorize_fast};
let gray = decolorize(&image); // fits a degree-2 polynomial mapping
let gray = decolorize_fast(&image); // picks the best of 66 channel weightingsdecolorize reproduces the color contrast most closely; decolorize_fast is
about an order of magnitude cheaper, because it searches a fixed number of pixel
pairs instead of solving for a mapping. Both take any image pixel type
(Rgb, Rgba, Luma, LumaA) with u8, u16, f32 or f64 samples, and
return Luma, or LumaA where the input carries alpha.
decolorize_with and decolorize_fast_with take an options struct if you need
to tune σ, the iteration budget or the sampling.
Enabling the default rayon feature parallelizes the work without changing the
output. Results are byte-identical either way.
Both variants are ports of the work of Cewu Lu, Li Xu and Jiaya Jia, whose project page collects the papers, the benchmark data and their reference implementation.
- Contrast Preserving Decolorization — ICCP 2012. The algorithm
behind
decolorize: a degree-2 polynomial color mapping whose nine coefficients are fitted by maximum likelihood under a bimodal Gaussian prior on the contrast of each neighboring pixel pair, solved by fixed point iteration. - Contrast Preserving Decolorization with Perception-Based Quality Metrics — IJCV 2014. The journal version, which adds the CCPR and E-score metrics used to evaluate this port.
- Real-time Contrast Preserving Decolorization — SIGGRAPH Asia 2012
Technical Briefs. The algorithm behind
decolorize_fast: the same idea with the weak color order dropped and the mapping restricted to convex combinations of the three channels, quantized to 66 candidates and scored on a sparse sample of pixel pairs.
The authors' own implementation ships in OpenCV as cv::decolor.
decolorize matches it to a mean absolute correlation of 0.992 across the
24 image benchmark, differing only in that it evaluates the published objective
for its convergence test rather than the smoothed variant OpenCV uses.