05.22.12 - Permalink
The information age we live in exists on top of a large amount of digital data. Thanks to the web we have access to this data. Thanks to search engines that index this data we can efficiently search it. What many may not realize is that this is only a fraction of the data that could be at our disposal. For millennia humans have recorded information on media such as paper. This information exists today in libraries and archives throughout the world with much of it not truly being a part of the information age.
Most text we find on the web was entered digitally by a human who typed it on a keyboard. To go back and do this for all the archived records created before the invention of the computer is an enormous and costly task. A much more realistic and practical means of digitizing this information is to scan it as a digital image. Digitizing old data in this manner, however, only solves one of our two problems, that of providing access to the data, not searchable access. This problem is exemplified by work being undertaken for the 1940s Census data release which will be in the form of 3.8 million JPEG images. Anyone can download this data and look at it. But then what? Can one find anything by looking through these millions of images containing billions of individual entries?
Though still a relatively young field, it is problems like these where the field of computer vision can offer some solutions. The burgeoning amounts of raw image data many organizations, agencies, and companies must deal with today have made even imperfect solutions highly desirable if they allow one to make some sense of large image collections. The Image, Spatial, and Data Analysis Group (ISDA), formed nearly a decade ago by Peter Bajcsy and currently led by Kenton McHenry, conducts research and development involving image and video data. Sifting through the state of the art in computer vision the ISDA group applies research to real world problems while striving to build robust multipurpose software to serve a variety of community needs. The group works on providing a framework for searchable access to large collections of handwritten forms, creating suites of semantic keyword extractors from image content, developing tools to aid in the study of group interactions in collections of videos, developing tools to aid in the identification of authorship in old manuscripts, and stitching together large macro images from collections of overlapping photographs. With decades of paper archives in the United States alone coupled with low-cost ubiquitous digital cameras that we all have today, efforts to unlock large collections of image data is becoming a necessity while also being very much a computationally intensive supercomputing problem.