Devices for scanning paper can be found in many work environments. Common examples range from the office photocopier to the fax machine. These two particular devices are similar in that the result of the scanning operation is traditionally printed on paper. This has the disadvantage that if a further copy of a document is required, then the original needs to be scanned again.
The increasing importance of the internet as a communication medium has recently led to a shift in focus within the document imaging industry. The typical destination for document images is no longer assumed to be hard copy. Increasingly an electronic version is required for storage in a document database that can be searched on-line. In turn, this has led to a rise in the importance of scanners.
Today's scanners typically come in one of three main forms: the flat-bed scanner, the sheet-feed (``keyboard'') scanner and the hand-held scanner. All three are rather cumbersome to use. The flat-bed and sheet-feed scanners require that the document be moved from its place of reading for scanning. The flat bed is worst in this respect as it takes up rather a lot of desk space and is therefore often situated away from the reader's desk. The sheet-feed scanner usually resides on the reader's desk but only takes single sheets of a limited range of paper sizes and therefore cannot handle newspapers, magazines, books or bound reports. Those documents it can scan typically need staples removing, and a time-consuming manual feed. The hand-held scanner usually requires several ``swipes'' of the scanning head to capture a desired region. This is inconvenient especially if the page surface is not flat (eg. bound books).
Over-the-desk scanning with video cameras [ 13 ] has many advantages over traditional scanning techniques. Perhaps the most obvious is that of convenience: documents need not be moved from their usual place of reading, so the user is able to scan without the usual overhead of moving the document. This promotes the act of ``casual'' scanning: the user is able to scan small amounts of interesting paper information as they are encountered whilst reading, rather than having to make a note of their position in a document so that they can be scanned or copied at a later stage [ 12 ]. In addition, the non-contact nature of the process facilitates the scanning of pages from bulky or fragile volumes.
Current standard video cameras enable full A4 pages to be sampled at about 50 greyscale dots per inch (dpi). This is not good enough for anything more than the crudest reproduction. While it is inevitable that higher resolution CCD cameras will, in due course, become commonplace and affordable, there is always going to be the need for even more resolution, for instance when scanning newspapers. This motivates the use of ``mosaicing'', whereby multiple, overlapping images of the document are stitched together seamlessly into one large, high resolution composite. Mosaicing techniques for general images abound in the literature [ 7 , 10 , etc,], and two calibrated cameras have been used to generate mosaics of documents [ 12 ]. Here we present a novel, uncalibrated, single-camera system for document mosaicing. What distinguishes this work is that we exploit knowledge of the application domain to develop a particularly robust and efficient mosaicing system.
A.H. Gee