Custom perceptual color distance for visually impaired people

Evaluating color differences is a crucial step in many computer vision applications. Most existing methods rely on perceptual color difference equations designed for observers with normal vision. However, these metrics are not tailored to visually impaired observers, whose color perception can differ substantially from that of the standard observer. In this paper, we propose a learning-based approach to estimate perceptual color differences adapted to observers with a specific visual impairment. We collect psychophysical data to train and evaluate our model and formulate the problem as a metric learning task. Since collecting large-scale psychophysical datasets is challenging, we introduce a lightweight Siamese neural network that combines limited human judgments with prior knowledge from conventional perceptual color distances. Experiments show that our architecture can predict perceptual color differences for unseen color pairs and generalizes across illumination conditions and to an independent dataset collected with new observers and stimuli. The proposed learned metric provides a flexible alternative for adapting color-based computer vision applications to specific visual impairments.