Spatio-Temporal Retinex-Inspired Envelopes with Anisotropic Diffusion

Since the introduction of the Retinex theory by Land and McCann in 1971, a multitude of different families, versions, interpretations, implementations, and applications have been proposed. The applications for image enhancement mainly differ in (i) how they explore the locality of the images to determine the local context, and (ii) how they recompute the pixel values based on this context. STRESS (spatio-temporal Retinex-inspired envelopes with stochastic sampling) is one of many quite successful members of the family of Retinex-based image enhancement algorithms. It explores the locality using a stochastic sampling technique, resulting in two envelope images - one maximum and one minimum envelope, completely enclosing the image signal and serving as a representation of the local image context. In this paper, we propose to exchange the stochastic sampling technique of STRESS, which causes significant chromatic noise, with an adapted version of constrained linear anisotropic diffusion for computing the envelopes, resulting in almost noise-free images. Using both subjective experiments and objective image metrics, we show that it improves the perceived and measured image quality and reduces noise artefacts. (C) 2023 Society for Imaging Science and Technology.