A machine learning approach to color space Euclidization

In this work, a machine learning methodology is proposed for the issue of color space Euclidization. Given a color difference formula as reference distance law, the Euclidization task consists in finding an injective transformation from the original color space into a real vector space and the corresponding inverse transformation, such that the Euclidean distances in the embedded color space align with the reference color distances. For this, artificial neural networks are devised as function approximators for the color space transformations being sought. Training these neural networks is accomplished through unsupervised learning, making use of random sampling and gradient descent. As key disagreement measure, either the (symmetric) relative isometric disagreement or the standardized residual sum of squares (STRESS) index is considered at a time and incorporated as part of the optimization criterion into the objective function. Comparative evaluation is carried out on well-established color distance laws, including the CIELAB-based DE2000 color difference formula. The evaluation results indicate significant performance advantages of the proposed approach over previous contributions.