Abstracts

Meirav Galun, Weizmann Institute Segmentation and denoising are important tasks in image analysis. One approach to these tasks is via the Segmentation by Weighted Aggregation algorithm (SWA). Inspired by algebraic multigrid (AMG), the SWA algorithm constructs in an adaptive way multiscale representations of the input image, useful for denoising and segmentation. In this talk we show that the coarsening approach of AMG is intimately related to anisotropic diffusion. Specifically, we show both theoretically and experimentally that the (two-level) algebraic multigrid restriction + prolongation operator implements a low-rank variation of anisotropic diffusion. We further compare its denoising properties to those of the classical anisotropic diffusion on smooth and step signals.

Irad Yavneh, Technion Multigrid algorithms for discretized nonlinear partial differential equations and systems are nearly as old as multigrid itself. Over the years several approaches and variants of nonlinear multigrid algorithms have been developed. Typically, for relatively easy problems the different approaches exhibit similar performance. However, for difficult problems the behavior varies, and it is not easy to predict which approach may prevail. In this talk we will consider nonlinear multigrid, focusing on the task of coarse-grid correction, in a general framework of variational coarsening. Such a view reveals clear relations between the various existing approaches and may suggest future variants. This study also sheds light on the choice of inter-rid transfer operators, which are so important for obtaining fast multigrid convergence, and which have received much attention in linear multigrid algorithms but far less so in nonlinear multigrid.

Raanan Fattal, Hebrew University Stochastic point distributions with blue-noise spectrum are used extensively in numerical analysis for function evaluation and computer graphics for various applications such as avoiding aliasing artifacts in ray tracing, halftoning, stippling, etc. In this talk we present a new approach for generating point sets with high-quality blue noise properties that formulates the problem as a statistical mechanics particle model and produces the points by sampling this model. This formulation unifies randomness with the requirement of equidistant point spacing, which leads to the enhanced blue noise spectral properties. We derive a highly efficient multi-scale sampling scheme to draw random point distributions from this model and avoid the critical slowing down phenomena that plagues this type of models. This derivation is accompanied by a model-specific analysis. Altogether, our approach generates high-quality point distributions, supports variable spatial point density, and runs in time that is linear in the number of points generated.

Boaz Nadler, Weizmann Institute The Non-local means algorithm is considered as one of the state-of-the-art image denoising algorithms. In this talk we'll first provide two different interpretations for non-local means: The first is a probabilistic interpretation in terms of a diffusion process in patch space. The second interpretation is as a proxy to the minimal mean squared error estimator within a Bayesian framework. Both interpretations have interesting consequences regarding the performance of NL-means and fundamental limitations for natural image denoising.

Nir Sochen, Tel Aviv University In some recent applications the objects of interest are fields of matrices. Diffusion Tensor MRI or in short DTI is an example of such application. Sets of matrices with the usual matrix multiplication are often Lie groups or quotients thereof. In this talk we will present variational regularization of such objects via anisotropic diffusion and the Beltrami framework. Applications to neuroimaging will be presented.

Michael Bronstein, Technion Large databases of 3D models available in public domain have created the demand for shape search and retrieval algorithms capable of finding similar shapes in the same way a search engine responds to text quires. Since many shapes manifest rich variability, shape retrieval is often required to be invariant to different classes of transformations and shape variations. One of the most challenging settings in the case of non-rigid shapes, in which the class of transformations may be very wide due to the capability of such shapes to bend and assume different forms. In this talk, we will explore approaches to 3D shape retrieval analogous to feature-based representations popular in the computer vision community. We will show how to construct invariant local feature descriptors based on heat diffusion in order to represent 3D shapes as collections of geometric "words" and "expressions" and how to adopt methods employed in search engines for efficient indexing and search of shapes. To conclude, we will show "Shape Google", a prototype search engine for deformable objects. (Based on joint work with M. Ovsjanikov, L. Guibas, A. Bronstein, I. Kokkinos, D. Raviv and R. Kimmel)

Ronen Basri, Weizmann Institute One of the most intensively studied problems in image processing concerns how to detect edges in images. Edges are important since they mark the locations of discontinuities indepth, surface orientation, or reflectance, and their detection can facilitate a variety of applications including image segmentation and object recognition. Accurate detection of faint, low-contrast edges in noisy images is challenging. Optimal detection of such edges can potentially be achieved if we use filters that match the shapes, lengths, and orientations of the sought edges. This however requires search in the space of continuous curves. In this talk we explore the limits of detectability, taking into account the lengths of edges and their combinatorics. We further construct two efficient multi-level algorithms for edge detection. The first algorithm uses a family of rectangular filters of variable lengths and orientations. The second algorithm uses a family of curved filters constructed through a dynamic-programming-like procedure using a modified beamlet transform. We demonstrate the power of these algorithms in applications to both noisy and natural images, showing state-of-the-art results.