Applications of Computing in Industry : Lecture

Generic Visual Categorization (GVC) refers to the association of human-defined categories, such as .chair, table, cat or dog., with photographic images, in general: that is, whatever the target category, imaging conditions or background content. This subject has seen explosive growth: a decade ago, computer vision researchers had only worked on a limited set of specific categories such as .face., .person., .military target. and .vehicle.; this goal was then viewed as a topic for psychologists and unrealistically ambitious for computer scientists; today, Google scholar identifies over 100,000 publications relating to .visual categorization..

The application we originally envisioned was to enable mobile phone users to find out more about the world around them simply by taking pictures. However, numerous unexpected applications were subsequently discovered, for instance: .content-based image enhancement. which decides to make images of flowers more colourful, while not making faces red; .personal photo print management. which controls the printing of large volumes of personal images on company devices; as well as more-expected applications to helping graphic designers find illustrations, tagging personal photo databases, categorizing document images and the contextual augmentation of images with advertising information.

First we give a brief overview of ongoing research at XRCE. We then review the basic machine learning techniques for visual categorization and the many applications thereof. Finally, we give a glimpse into the implications of the simple observation that .similarity. . on which most machine learning techniques are based . is itself a contextually-dependent notion.

Christopher Dance, Research Fellow, Xerox Research Centre Europe (XRCE), Grenoble, France, http://www.xrce.xerox.com

Christopher received a BA (Physics and Theoretical Physics) and PhD (Information Engineering) from Cambridge University, England, where he worked on machine learning techniques for classifying quark jets and medical diagnosis, as well as geometrical techniques for reconstructing three-dimensional models from multi-axial cross-sections. He joined XRCE in 1997, where he conducted research and technology transfer relating to mobile imaging, character recognition, user interfaces and visual categorization. He set up XRCE's textual and visual pattern analysis (TVPA) and machine learning for optimization and services (MLS) research groups in 2002 and 2006, and was XRCE's Laboratory Manager from 2004-2009. More recently his research interests have included coupling machine learning with stochastic optimization and game theoretic methods for services and supply chains. He has published over 30 refereed academic papers and has been granted over 20 patents.