Xerox scientists develop system for categorizing 'generic' electronic images
NOVEMBER 17--Scientists at the Xerox Corporation Research Centre in Europe have developed a powerful new system for "recognizing" generic, everyday objects in digital images, such as a photograph of a car, and categorizing them.
NOVEMBER 17--Scientists at the Xerox Corporation Research Centre (XRCE) in Europe have developed a powerful new system for "recognizing" generic, everyday objects in digital images, such as a photograph of a car, and categorizing them. When used in document and content-management systems, this breakthrough technology would allow people to filter and search for images, as well as text. It would enable efficient storage and management of electronic images and could significantly extend Web-searching capabilities, which are currently based upon text only.
"Although there has been phenomenal growth in the use of digital cameras and images, the use of technology to categorize image content is in its infancy," said Christopher Dance, senior scientist and image-processing manager at Xerox Research Centre Europe in Grenoble. "It is currently only used in applications such as face recognition in the security industry."
However, scientists at the XRCE developed a generic technique for the identification of images, allowing the categorization of everyday image content such as buildings, animals, airplanes, books, and faces. It is a generic image-categorization technology that is robust, fast, and simple to use. The technology, which results from fundamental research at XRCE, melds the lab's expertise in image processing, computer vision, and machine learning.
Image categorization is analogous to text categorization, which looks at the content of a document to find key words. To categorize images, Xerox identifies the key features of an object, which it calls "patches." The system works by training a computer to map the patches and to classify sets of these patches. This classification in effect assigns an image to a particular category or categories. The scientists had to solve some knotty problems, Dance said. For example, early versions of the system could confuse an image of a stack of tires and an image of a car, as they both contain some of the same patches. To overcome this, the program examines key patches in the context of other areas of the picture. In this example, a stack of tires would not get confused with a car because the machine would recognize other key patches such as headlights or windows were missing.
"Images play a key role in most documents, but in the past document repositories have only been able to search for and categorize text," Dance said. "We will be working with Xerox business groups to integrate this new system into Xerox's document-management offerings, providing customers with an additional competitive advantage."
In addition to developing this software for different applications, Xerox will continue to extend its categorizer to handle more visual categories and to incorporate difficult cases where the object of interest occupies only a small fraction of the field of view, Dance said.