This page contains some example stats obtained by using Classnotes. When starting a new profile the stats should help give some idea as to how much training is necessary to have a high per letter accuracy when parsing new scans. In general, a handwriting profile will require significantly more instances per character to be trained than a font profile. Ideally, the program should be trained with a sufficient number of characters so that the per letter accuracy in translating a new scan will be in the 90% and up range. If this accuracy percentage is reached it should be trivial to perform an automated spell check on the text produced from a new scan and have it yield high per word accuracy.
If any user has an interesting profile send me an e-mail with the profile and a sample scan not used in training. I will take a look at it and possibly post it to this page.
Sarah UPPERCASE:
This profile has been trained with a sample of all CAPS handwriting. It has the following stats:
- Avg trained instances per character: 66.7 (min: 15; max: 187)
- Per letter accuracy in new scans: 69.4 %
Action Man UPPERCASE:
This profile has been trained with a sample of all CAPS characters using the Action Man font. It has the following stats:
- Avg trained instances per character: 4.69 (min: 8; max: 15) (includes the unused lowercase letters)
- Per letter accuracy in new scans: 100 %
Files: [Profile] [Sample Scan]
Addjazz UPPERCASE:
This profile has been trained with a sample of all CAPS characters using the Addjazz font. It has the following stats:
- Avg trained instances per character: 4.69 (min: 8; max: 15) (includes the unused lowercase letters)
- Per letter accuracy in new scans: 100 %
Files: [Profile] [Sample Scan]
