Dikla Horesh (EE, Technion)
Tuesday, 14.7.2015, 11:30
The total variation (TV) functional is today a fundamental regularizing tool in image processing. It is employed for denoising, deconvolution, optical-flow, tomographic reconstruction, texture and image analysis and more. Recently a spectral TV framework was introduced which extends linear filtering techniques and eigenvalue analysis to a convex nonlinear setting, and specifically to the TV functional.
In this work we use the Spectral TV Domain to design advanced texture decomposition and processing algorithms. The first part of the work aims at obtaining precise local texture orientations of images in a multi-scale manner. We show that using this method one can detect and differentiate a mixture of overlapping textures and obtain with high fidelity a multi-valued orientation representation of the image.
In the second part, we introduce a novel notion of a separation surface. It is a separation map locally varying in the Spectral TV Domain. It can be used for decomposition of textures with spatially varying scale or illumination. A texture processing application is proposed where selected textures can be easily either enhanced or attenuated, according to user control, with a very realistic appearance.
* This talk summarizes the M.Sc. research under the supervision of Prof. Guy Gilboa