Not surprisingly, recognition of pre-printed numerics was far superior to that of hand-written numerics. At comparable rejection rates of around 12%, for the RE engine with image enhancement, the digit substitution rate for hand-written numerics was 1.41% compared to the rate of 0.16% for pre-printed numerics.

For the RE engine on enhanced images, we inspected all 43 cases for 'district' or 'subdistrict' where the keyed and recognised responses differed, and the recognised response would be automatically accepted when the confidence threshold for all digits was 650. In none of the cases was the response on the image pre-printed and whole.

As can be seen from the table above, in 40% of the cases viewed the 'district' or 'subdistrict' field had actually been filled out by hand. In 56% of the cases viewed the pre-printed 'district' or 'subdistrict' had the tops of the characters cut off. This only occurred on individual form fronts (IF's) where, for some reason, the printing in the top right hand corner of the form had been cut from the image. Unfortunately, we presume some sort of registration error caused part of the 'district' and 'subdistrict' responses to be cut out also.

For the RE engine on enhanced images, we never found an example of a whole pre-printed digit being substituted for another digit.

Rejection rates for pre-printed numerics were higher than could be expected. For the RE engine on enhanced images when the confidence threshold for all digits was 0, 11% of data was still rejected. Upon discovering the cutting error referred to above, we decided to analyse the rejection rates separately for DF's and IF's (expecting the rates to be worse on IF's).

Surprisingly the rejection rates were worse for the DF's. For DF's, if the 'subdistrict' had been rejected, there was an 86% chance that the 'district' had also been rejected. Of the few cases viewed, the 'district' and 'subdistrict' were clearly pre-printed. So why were the rejection rates so high? We suspect that this relates back to the original difficulty that both engines had in registering the DF's at all.

For the RE engine on enhanced images, we also looked at the proportion of rejected 'district' and 'subdistrict' responses that were rejected only because they contained an excessive number of digits, and not because one or more of the digits had a confidence score below 650.

There were 228 recognised 'district' and 'subdistrict' responses rejected only because they were too long, where the keyed response was available for comparison. All of these were too long because they contained one or more recognised hyphens, "-". Had the hyphen not been there, every recognised response would have been correct. That is, the recognised digits were all correct, but a hyphen had been inserted somewhere in the recognised response. The recognition library should not have included a hyphen as a valid numeric character. Presumably different libraries were used for hand-written and pre-printed numerics because hyphens only featured as characters in the recognised 'district' and 'subdistrict' responses.