Poll

The ITinCVPR book ...
 
Foreword Print
Wednesday, 15 July 2009 17:23

Foreword (by Alan Yuille)

 

Computer vision and pattern recognition are extremely important research fields with an enormous range of applications. They are also extremely difficult. This may seem paradoxical since humans can easily interpret images and detect spatial patterns. But this apparent ease is misleading because neuroscience shows that humans devote a large part of their brain, possibly up to 50% of the cortex, to processing images and interpreting them. The difficulties of these problems have been appreciated over the last 30 years as researchers have struggled to develop computer algorithms for performing vision and pattern recognition tasks. Although these problems are not yet completely solved, it is becoming clear that the final theory will depend heavily on probabilistic techniques and the use of concepts from information theory.

The connections between information theory and computer vision have long been appreciated. Vision can be considered to be a decoding problem where the encoding of the information is performed by the physics of the world – by light rays striking objects and being reflected to cameras or eyes. Ideal observer theories were pioneered by scientists such as Horace Barlow to compute the amount of information available in the visual stimuli, and to see how efficient humans are at exploiting it. But despite the application of information theory to specific visual tasks, there has been no attempt to bring all this work together into a clear conceptual framework.

This book fills the gap by describing how probability and information theory can be used to address computer vision and pattern recognition problems. The authors have developed information theory tools side by side with vision and pattern recognition tasks. They have characterized these tools into four classes: (i) measures, (ii) principles, (iii) theories, and (iv) algorithms. The book is organized into chapters addressing computer vision and pattern recognition tasks at increasing levels of complexity. The authors have devoted chapters to feature detection and spatial grouping, image segmentation, matching, clustering, feature selection, and classifier design. As the authors address these topics, they gradually introduce techniques from information theory. These include (1) information theoretic measures, such as entropy and Chernoff information, to evaluate image features; (2) mutual information as a criteria for matching problems (Viola and Wells 1997); (3) minimal description length ideas (Risannen 1978) and their application to image segmentation (Zhu and Yuille 1996); (4) independent component analysis (Bell and Sejnowski 1995) and its use for feature extraction; (5) the use of rate distortion theory for clustering algorithms; (6) the method of types (Cover and Thomas 1991) and its application to analyze the convergence rates of vision algorithms (Coughlan and Yuille 2002); and (7) how entropy and infomax principles (Linsker 1988) can be used for classifier design. In addition; the book covers alternative information theory measures, such as Rényi alpha-entropy and Jensen–Shannon divergence, and advanced topics; such as data driven Markov Chain Monte Carlo (Tu and Zhu 2002) and information geo-
metry (Amari 1985). The book describes these theories clearly, giving many illustrations and specifying the code by flow -charts.

Overall, the book is a very worthwhile addition to the computer vision and pattern recognition literature. The authors have given an advanced introduction to techniques from probability and information theory and their application to vision and pattern recognition tasks. More importantly, they have described a novel perspective that will be of growing importance over time. As computer vision and pattern recognition develop, the details of these theories will change, but the underlying concepts will remain the same.

 

Alan Yuille

UCLA, Department of Statistics and Psychology
Los Angeles, CA
March 2009

 

 

(© 2009, Springer)

Last Updated on Thursday, 30 July 2009 10:17