Direction-depedent variational models for image processing and vision
vendredi 1 février 2019, 14h00 - 16h00
The use of anisotropic approaches has a long tradition in the imaging community. Historically, such methods date back to the seminal work of J. Weickert where anisotropic non-linear diffusion PDEs were applied to solve several image restoration problems such as denoising, inpainting etc.
Recently, direction-dependent models has been reinterpreted in variational scenarios and applied to a wide variety of imaging tasks due to their better adaptation to describe image statistics and, consequently, their better performance in image reconstruction problems.
In this talk, I will first present an anisotropic variant of the transport-diffusion osmosis model applied to solve light-balance and image fusion problems and present a new image regulariser which generalises the standard TV prior to better describe natural image statistics.
In the second part of the talk, I will draw connections between classical direction-dependent approaches and my ongoing work on cortical-inspired contrast and color perception models. Here, the local orientation appears as an independent variable as a result of a `lifting’ procedure which extends the image from a bi- to a three-dimensional space so as to model the sensitivity to local orientation of neurons in the primary visual cortex (V1).