Interpolation for nonlinear Retinex-type algorithms

In this paper we propose a method to speed up Retinex-type algorithms, consisting of a computationally intensive nonlinear illumination estimation module followed by a relatively simple manipulation module. Speed up is obtained by way of computing the illumination on a sub-sampled image. The challenge is to interpolate a piece-wise smooth low resolution image. We present and analyze the trade off between two types of interpolation methods. On one hand, regular illumination interpolation, which preserves the Retinex-type output quality, however may result in artifacts. On the other hand a detail preserving interpolation which removes artifacts, however may compromise the output quality.