A computational method for predicting color palette discriminability

Automatic analysis of images is increasingly being used to generate color insights and this has led to various methods for generating palettes. Several studies have recently been published that explore methods to predict the visual similarity between pairs of palettes and these methods are often used to evaluate different generative methods. This work is concerned with being able to predict visual similarity between color palettes. Three data sets (two of which were previously published) are used to evaluate two methods for predicting visual similarity between palettes. A novel palette-difference metric (based on the Hungarian algorithm) is compared to the previously published minimum color difference model (MICD) and was found to agree better with the visual data for two of the three data sets. Agreement between models and visual data was also better for CIEDE2000 (1, 2) than for CIELAB metrics.