XI Workshop de Visão Computacional
October 05th - 07th 2015 - São Carlos (SP) - Brazil

Keynote Speakers

"We are sorry to inform that due to personal reasons,
Prof. Yannis Aloimonos and Prof. Mounim Yacoubi
will not be able to attend the workshop this year"

 

Patch Foveation in Nonlocal Image Filtering

Alessandro Foi, Ph.D.
Department of Signal Processing
Tampere University of Technology, Tampere, Finland
http://www.cs.tut.fi/~foi/

Short Bio: Alessandro Foi received the M.Sc. degree in Mathematics from the Università degli Studi di Milano, Italy, in 2001, the Ph.D. degree in Mathematics from the Politecnico di Milano in 2005, and the D.Sc.Tech. degree in Signal Processing from Tampere University of Technology, Finland, in 2007. He is currently an Academy Research Fellow with the Academy of Finland, at the Department of Signal Processing, Tampere University of Technology, where he is also Associate Professor. His research interests include mathematical and statistical methods for signal processing, functional and harmonic analysis, and computational modeling of the human visual system. His recent work focuses on spatially adaptive (anisotropic, nonlocal) algorithms for the restoration and enhancement of digital images, on noise modeling for imaging devices, and on the optimal design of statistical transformations for the stabilization, normalization, and analysis of random data. He is a Senior Member of the IEEE, Member of the Image, Video, and Multidimensional Signal Processing Technical Committee of the IEEE Signal Processing Society, and an Associate Editor for the IEEE Transactions on Image Processing and for the new IEEE Transactions on Computational Imaging.

Abstract: When we gaze a scene, our visual acuity is maximal at the fixation point (imaged by the fovea, the central part of the retina) and decreases rapidly towards the periphery of the visual field. This phenomenon is known as foveation. To form a complete image of the scene, the human visual system (HVS) typically processes a multitude of foveated retinal images gathered at different fixation points. In this talk we look at the analogies and connections between this feature of the HVS and modern nonlocal (NL) image filters. NL filters rely on the assumption that natural images contain a large number of mutually similar patches at different locations within the image: similar patches are first identified, and then used into adaptive weighted averages or more sophisticated nonlinear shrinkage. Such approach is at the core of several of the most effective image restoration methods to date. Crucial elements in the design of NL filters are the metric or distance used for assessing the patch similarity, and the size of the patch. Large patches guarantee stability of the distance with respect to degradations such as noise; however, the mutual similarity between pairs of patches typically decreases as the patch size grows. Thus, a windowed Euclidean distance is commonly employed to balance these two conflicting aspects, assigning lower weights to pixels far from the patch center. Choosing a metric for patch similarity corresponds to assuming a specific model for describing natural images and their self-similarity: the effectiveness of NL methods depends strongly on the validity of such underlying model. We particularly investigate a different form of self-similarity: the foveated self-similarity. Foveation here corresponds to a spatially variant blur operator, characterized by blur kernels whose bandwidth decreases with the spatial distance from the patch center. In contrast with the conventional windowing, which is only spatially selective and attenuates sharp details and smooth areas in equal way, patch foveation provides selectivity in both space and frequency, mimicking the HVS inability to perceive details at the periphery of the center of attention. Throughout the talk, we adopt the image denoising problem as a simple means of assessing the effectiveness of descriptive models for natural images. We show that, in nonlocal image filtering, the foveated self-similarity is far more effective than the conventional windowed self-similarity. To facilitate the use of foveation in nonlocal imaging, we present a general framework for designing foveation operators, i.e. linear operators producing foveated patches by means of spatially variant blur. Within this framework, several parametrized families of foveation operators are demonstrated, including anisotropic ones. Strikingly, the operators enabling the best denoising performance on complex natural images are the radial ones, in complete agreement with the orientation preference of the HVS.

 

Computer Vision in Medical Imaging and Measurements

Jacob Scharcanski, Ph.D.
UFRGS - Universidade Federal do Rio Grande do Sul
Instituto de Informática, Rio Grande do Sul, Brasil
http://www.inf.ufrgs.br/~jacobs/

Short Bio: Jacob Scharcanski is a (Full) Professor in Computer Science at the Federal University of Rio Grande do Sul (UFRGS), Brasil. He holds a cross appointment with the Department of Electrical Engineering at UFRGS, and also is an Adjunct Professor with the Department of Systems Design Engineering, University of Waterloo, Canada. He authored and co-authored over 150 refereed journal and conference papers, book chapters and books, and delivered over 30 invited presentations worldwide. He serves as an Associate Editor for two journals, and has served on dozens of International Conference Committees. In addition to his academic activities, he has several technology transfers to the private sector. Professor Scharcanski is a licensed Professional Engineer (PEO, Canada), Senior Member of the IEEE, Member of SPIE, and serves as Co-Chair of the Technical Committee IEEE IMS TC-17 (Imaging Measurements and Systems).

Abstract: In this talk, computer vision in medical imaging and measurements is proposed as a way to facilitate the interpretation of phenomena based on medical imagery, or to make inferences based on models of such phenomena. In order to illustrate this presentation, several modeling issues in medical imaging and measurements are discussed, and illustrated by examples. When modeling imaging measurements, usually we are trying to describe the world (or a real world phenomenon) using one or more images, and reconstruct some of its properties based on imagery data (like shape, texture or color). Actually, this is an ill-posed problem that humans can learn to solve effortlessly, but computer algorithms often are prone to errors. Nevertheless, in some cases computers can surpass humans and help interpret imagery more accurately, given the proper choice of models, as we will discuss in this talk. Modeling medical imaging measurements often involves errors, and estimating the expected error of a model can be important in some applications (e.g. when estimating a tumor size and its potential growth, or shrinkage, in response to treatment). Typically, a model has tuning parameters, and these tuning parameters may change the model complexity. We wish to minimize modeling errors and the model complexity, in other words, to get the ‘big picture’ we often sacrifice some of the small details. For example, estimating tumor growth (or shrinkage) in response to treatment requires modeling the tumor shape and size, which can be challenging for real tumors, and simplified models may be justifiable if the predictions obtained are informative (e.g. to evaluate the treatment effectiveness). This issue is closely related to machine learning and pattern recognition, and techniques of these areas can be adapted to resolve problems in medical imaging measurements. To conclude this talk, open problems in medical imaging measurements and model selection are discussed in some detail.

 

Using Computer Vision with drones in Agriculture

Lúcio André de Castro Jorge, Ph.D.
EMBRAPA - Empresa Brasileira de Pesquisa Agropecuária.
São Carlos, Brasil.
http://lattes.cnpq.br/3036476562950521

Short Bio: Degree in Electrical Engineering - Electronics and Electrical Engineering at the School of Barretos Engineering (1987); Master's degree in Computational Mathematics and Computer Science at Institute of Mathematics and Computer Sciences at the University of São Paulo, ICMC-USP (2001); PhD in Signal Processing and Instrumentation from the School of Engineering, University of São Paulo, SEL-EESC-USP (2011); LatoSensu in image processing from the University of Campinas - Unicamp (1990); LatoSensu in Geografical Information Systems from the Federal University of São Carlos - UFSCar (2005); Researcher at Embrapa Instrumentation since 1990; Professor of Image Porcessing, Computer Graphics and Artificial Intelligence at UNISEB- Colleges COC since 2006. Experience in Computer Science, working on the development of image processing softwares, embedded systems, mobile devices (PDAs), pattern recognition and intelligence computing, computer graphics and geo-referenced systems. Experience applied in several projects in Agriculture, Precision Agriculture, GIS, agricultural monitoring, remote sensing, study of roots, leaves, plant diseases and deficiencies, development of UAV (unmanned aerial vehicle) for agricultural use.

Abstract: The use of drones in agriculture increase the use of advanced sensors to evaluate the development of different crops providing additional agricultural knowledge to deal with the variability in the field in order to contribute to increased crop yields. The technological development of remote sensing techniques using drones or UAVs, have been improving the crop management with greater spatial and temporal resolution. Therefore, the main challenge in this area is the development of methods of processing images quickly and accurately working with large volumes of multidimensional and temporal datas. However, despite the enormous potential of this remote sensing, their implementation in practice is very limited. This work will be presented the state of the art at all stages, from image acquisition, processing, segmentation, classification among others for applications in agriculture.