Abstract:
Despite the remarkable progress achieved by modern neural networks, the problems of color registration and modeling remain unsolved. In particular, the dominant approaches for color coordinate transformation, especially from sensor color space, still rely on low-parameter regression methods, such as polynomial and root-polynomial linear regression. This is primarily due to extremely stringent computational efficiency requirements. Nevertheless, even within this narrow scientific niche, fundamental questions persist: what do we actually know about optimal color transformation? Which properties must it satisfy and which properties are inherently impossible for it to possess? In attempts to advance toward solving this problem, a research group comprising K. Soshin, M. K. Chobanu, D. P. Nikolaev, and E. I. Ershov proposed a new scalable rational transformation, while A. Shutova and I. Ermakov developed mathematical estimates for the asymptotic growth rate of the feature space. The upcoming talk will outline the development path, key findings, and remaining open questions in this area.
|
