The Faculty of Informatics is pleased to announce a seminar given by Giacomo Boracchi
 
TITLE: Foveated self-similarity in nonlocal image filtering
SPEAKER: Giacomo Boracchi, Universita Statale degli Studi di Milano
DATE: Tuesday, September 24, 2013
PLACE: University of Lugano, room A24, Red building (Via G. Buffi 13)
TIME: 11.30

ABSTRACT:
Nonlocal self-similarity is widely acknowledged as an effective regularization prior for natural images. To exploit the nonlocal self-similarity, similar patches have to be identified on the basis of a suitable patch distance: choosing the patch distance actually implies the assumption of a specific descriptive model for natural images and the effectiveness of many algorithms depend on the reliability of such underlying model. Patch similarity is typically assessed through a windowed Euclidean distance of the pixel intensities.
 
Inspired by the human visual system, we design specific foveation operators to measure the patch similarity. Foveation is a peculiarity of the HVS, which is characterized by maximal acuity at the fixation point (imaged by the fovea) that decreases towards the periphery of the visual field. To reproduce such effect, foveation operators yield patches blurred by a spatially variant blur, where the point-spread functions (PSFs) have standard deviation that increases with the spatial distance from the patch center.
 
We measure patch similarity by means of the foveated distance, i.e., the Euclidean distance between foveated patches, and we investigate the foveated self-similarity. To quantitatively validate the foveated self-similarity as regularization prior for natural images, we consider the image denoising problem and specifically the Nonlocal Means (NL-means) for the removal of additive white Gaussian noise. We introduce Foveated NL-means, where the foveated distance replaces the conventional windowed distance, and we run extensive tests to show that Foveated NL-means can substantially outperform the standard one.
 
We motivate the superior performance of Foveated NL-means in terms of low level vision: indeed, the denoising performance can be treated as a compact indicator of the ability to identify similar patches and to distinguish between different ones in noisy environments. This makes the foveated self-similarity a far more effective regularization prior for natural images than the conventional windowed self-similarity.
 
Joint work with Alessandro Foi (Tampere University of Technology)
 
 
BIO:
Giacomo Boracchi received the M.S. degree in Mathematics from the Universita Statale degli Studi di Milano, Italy, and the Ph.D. degree in Information Technology at Politecnico di Milano, Italy, in 2004 and 2008, respectively. He was visiting Tampere International Center for Signal Processing, Finland, during 2004-2005.
Currently, he is a postdoctoral researcher at the Dipartimento di Elettronica, Informazione e Bioingegneria of the Politecnico di Milano. His main research interests focus on mathematical and statistical methods for image/video processing and for designing intelligent systems operating in nonstationary environments.
 
 
HOST: Dr. Davide Eynard