A new method of reading the handwritten zip codes in the U.S. Postal Services CD-ROM database is presented. Zip code images are binarized, segmented and recognised. A recognition driven method for splitting multiple connected digits has been developed; for grouping together of broken digits, the system targets components with near-touching stroke tips, 5-hats, and 4-Ls. The digit recogniser is a majority vote combination of 3 neural networks with a zero rejection performance of 96.53% on the 2711 imperfectly segmented digits in the cedarbs test set. With digit splitting capability disabled, the system performance on the 930 whole zip codes of the test set is 61.0% correct with no errors when up to two rejected symbol positions are allowed. With digit splitting enabled the performance rises to 66.3%.
A new system for reading handwritten zip codes
1995-01-01
436777 byte
Conference paper
Electronic Resource
English
A New System for Reading Handwritten Zip Codes
British Library Conference Proceedings | 1995
|Reading handwritten US census forms
IEEE | 1995
|Reading Handwritten US Census Forms
British Library Conference Proceedings | 1995
|A Probabilistic Approach to Automatic Handwritten Address Reading
British Library Conference Proceedings | 1995
|