Title:
Object Description: Accuracy, Noise, and Applications
Date:
Friday, October 22, 2010 from 2:00pm – 3:00pm
Coffee and light refreshments are served in E2-361 from 1:30 P.M.
Location:
E3-262 EITC Building (Senate Chambers), University of Manitoba Fort Garry Campus
Speaker:
Dr. Miroslaw Pawlak
Department of Electrical and Computer Engineering
University of Manitoba
Abstract:
This talk is about the problem of image characterization and analysis via the domain-based descriptors utilizing the mathematical theory of moments and orthogonal polynomial bases. The moments and orthogonal polynomials are classical concepts in mathematical analysis and statistics not only due to their simplicity and own elegance but also for the extraordinary range of subjects and applications where they have illuminated. In image analysis they can be directly applied to invariant pattern recognition, image normalization, image registration, image symmetry, and watermarking.
In this talk we focus on a specific class of image descriptors stemming from the theory of moments and orthogonal polynomials of two variables as well as the theory of invariance. The fundamental accuracy analysis of the introduced image descriptors is given. This includes assessing the performance of the descriptors with respect to common data deformations such as discretization and noise. We also examine an error due to the geometric nature of the image plane and this is referred to as the geometric error. This type of error is explained by finding the connection between the accuracy issue with the analytic number theory of lattice point approximations. Several new techniques to increase the accuracy and efficiency of moment descriptors are also proposed. We utilize these results for solving the problem of reconstruction of noisy images from orthogonal moments. The theory developed reveals fundamental trade-offs between the aforementioned types of errors. This leads to the issue of an intrinsic dimensionality of a feature vector which yields an optimal image representation.
The obtained fundamental results are employed to tackle two important problems of image analysis. The first one concerns symmetry detection. Detection of symmetries and symmetry-based representations constitute important practical issues that have not received much attention in the image analysis literature. We develop formal statistical procedures to test symmetries existing in an image observed in the presence of noise. In the second application we propose a novel watermarking system able to cope with geometric transformations. Watermarking algorithms form a critical part of modern multimedia systems where data hiding and copyright protection issues are very essential.
Speaker Bio:
Miroslaw Pawlak received the Ph.D. and D.Sc. (habilitation) degrees in computer engineering from Wroclaw University of Technology, Wroclaw, Poland.
He held research and teaching positions at Wroclaw University of Technology and Concordia University, Montreal. He is currently a Professor at the Department of Electrical and Computer Engineering, University of Manitoba. He has held a number of visiting positions in North American, Australian, and European Universities. He was at the University of Ulm and Georg-August University in Goettingen, Germany as the Alexander von Humboldt Foundation Fellow.
His research interests include statistical aspects of signal/image processing, machine learning, and nonparametric modeling. Among his publications in these areas are the books Image Analysis by Moments: Reconstruction and Computational Aspects (WUT Press, 2006), and Nonparametric System Identification (Cambridge University Press, 2008), coauthored with Prof. Greblicki (Dr. Pawlak Ph.D. supervisor). His publications have appeared in various IEEE journals (IEEE: Information Theory, Pattern Analysis and Machine Intelligence, Image Processing, Signal Processing, Communications, Control, Circuit and Systems) as well as in statistical journals (Annals of Statistics, Journal of Multivariate Analysis, Annals of Applied Statistics).
Dr. Pawlak has been an Associate Editor of the Journal of Pattern Recognition and Applications, Pattern Recognition, International Journal on Sampling Theory in Signal and Image Processing, and Opuscula Mathematica.
Cost:
This will be a free event.
Contact:
For questions or more information: Jun Cai 474-6419
