Abstract
Recoloring (daltonization) helps observers with color-vision deficiency (CVD) distinguish colors they would otherwise confuse, but a single global transform changes the whole image and tends to look unnatural. We propose targeted recoloring, which concentrates the color change where it is needed. A learned lost-saliency map, the positive part of the difference between normal-vision and dichromat saliency, marks the regions that are salient in normal vision but lose attention under dichromacy, and recoloring is concentrated on those regions by weighting the optimization with that map, in CIELAB with lightness fixed and only the chromatic channels free. The dichromat saliency needed to build the map is predicted from about 100 images by adapting a frozen DINOv2 backbone with low-rank adaptation. On the EToCVD benchmark, measured by predictor-independent color metrics, targeted recoloring improves both naturalness and distinguishability over full daltonization and a simplified Kuhn-style baseline, with the color change concentrated in the target region. Because the benchmark provides only fixations, a saliency-map target has to be made by smoothing, and we show that a fixed-pixel Gaussian on variable-resolution images is mis-calibrated and can reverse predictor rankings; we demonstrate this on the independent Toronto benchmark and remove it by calibrating to one degree of visual angle. The learned predictors are trained out of fold, and we report the limits of the predictor honestly.
Funding Information
  • Japan Society for the Promotion of Science (JP26K15753)