Training Dataset

Because of the geometric nature of the features, a more heterogeneous training dataset was needed. The concept exploration dataset lacked several font and pitch combinations; the features proposed can be affected by size variations. The training dataset was constructed with the following characteristics:

• Twenty four pages total, with 12 ``good'' and 12 ``bad'', where the meaning for ``good'' is an OCR median accuracy of at least 90%.

• Three pages were re-used from the concept exploration dataset and 21 new ones were selected from ISRI's ``Sample 2'' database.

• The pages were selected based on their median OCR accuracy (see below), font type and pitch. All the combinations were constructed and 3 pages were selected for each combination (see Table 3.3). The median accuracy was computed from the output of eight OCR devices (see Table 4.1) except for pages 2002-011, 5207-005 and 5319-008 that were processed by the devices listed in Table 3.2.

• Whenever possible, pages containing at least 500 characters were selected.

• Only text zones were considered. Tables and graphs were ignored.

The reason for using median accuracy instead of the mean accuracy is that the median measure is a more stable metric, since it is not affected by abrupt lows or highs in accuracy for any device. The mean value, on the other hand, would be affected by such a behaviour and thus would render an accuracy value that is not representative of the ``general'' accuracy OCR devices have on the page.