BOOSTING PERFORMANCE IN NEURAL NETWORKS

1993 ◽  
Vol 07 (04) ◽  
pp. 705-719 ◽  
Author(s):  
HARRIS DRUCKER ◽  
ROBERT SCHAPIRE ◽  
PATRICE SIMARD
Keyword(s):  
Neural Networks ◽  
Error Rate ◽  
Character Recognition ◽  
Optical Character Recognition ◽  
The United States ◽  
Pac Learning ◽  
Postal Service ◽  
United States Postal Service ◽  
Lower Case ◽  
Improved Performance

A boosting algorithm, based on the probably approximately correct (PAC) learning model is used to construct an ensemble of neural networks that significantly improves performance (compared to a single network) in optical character recognition (OCR) problems. The effect of boosting is reported on four handwritten image databases consisting of 12000 digits from segmented ZIP Codes from the United States Postal Service and the following from the National Institute of Standards and Technology: 220000 digits, 45000 upper case letters, and 45000 lower case letters. We use two performance measures: the raw error rate (no rejects) and the reject rate required to achieve a 1% error rate on the patterns not rejected. Boosting improved performance significantly, and, in some cases, dramatically.

Kurzweil Reading Machine: A Partial Evaluation of Its Optical Character Recognition Error Rate

1979 ◽  
Vol 73 (10) ◽  
pp. 389-399
Author(s):  
Gregory L. Goodrich ◽  
Richard R. Bennett ◽  
William R. De L'aune ◽  
Harvey Lauer ◽  
Leonard Mowinski
Keyword(s):  
Error Rate ◽  
Character Recognition ◽  
Optical Character Recognition ◽  
Partial Evaluation ◽  
Error Rates ◽  
Recognition Error ◽  
Optical Character ◽  
Printed Materials ◽  
High Level

This study was designed to assess the Kurzweil Reading Machine's ability to read three different type styles produced by five different means. The results indicate that the Kurzweil Reading Machines tested have different error rates depending upon the means of producing the copy and upon the type style used; there was a significant interaction between copy method and type style. The interaction indicates that some type styles are better read when the copy is made by one means rather than another. Error rates varied between less than one percent and more than twenty percent. In general, the user will find that high quality printed materials will be read with a relatively high level of accuracy, but as the quality of the material decreases, the number of errors made by the machine also increases. As this error rate increases, the user will find it increasingly difficult to understand the spoken output.

APGAR STAMP

PEDIATRICS ◽  
1994 ◽  
Vol 94 (4) ◽  
pp. 544-544
Author(s):  
L. J. Butterfield
Keyword(s):  
United States ◽  
Annual Meeting ◽  
San Antonio ◽  
The United States ◽  
Press Conference ◽  
Postal Service ◽  
United States Postal Service ◽  
American Academy Of Pediatrics ◽  
One Year ◽  
United States Postal

On Monday, October 24, 1994 at 2:00 PM, a definitive stamp will be dedicated to Dr Virginia Apgar at the American Academy of Pediatrics (AAP) annual meeting in Dallas. A definitive stamp lasts for years while the commemorative stamp is printed just one year. The United States Postal Service announced the 1994 stamp program on December 7, 1993 during a press conference at the National Postal Museum. Dr Apgar was nominated for a stamp in 1987 by the AAP. The initiative was spawned by the Perinatal Section at the 1985 annual meeting of the AAP in San Antonio.

Staff scheduling at the United States Postal Service

2003 ◽  
Vol 30 (5) ◽  
pp. 745-771 ◽  
Author(s):  
Jonathan F. Bard ◽  
Canan Binici ◽  
Anura H. deSilva
Keyword(s):  
United States ◽  
The United States ◽  
Postal Service ◽  
Staff Scheduling ◽  
United States Postal Service ◽  
United States Postal

Optical Character Recognition of Indian Language Manuscripts using Convolutional Neural Networks

2021 ◽  
pp. 894-911
Author(s):  
Bhavesh Kataria, Dr. Harikrishna B. Jethva
Keyword(s):  
Neural Networks ◽  
Character Recognition ◽  
Optical Character Recognition ◽  
Short Term Memory ◽  
Short Term ◽  
Indian Language ◽  
Term Memory ◽  
Text Document ◽  
Optical Character ◽  
Long Short Term Memory

India's constitution has 22 languages written in 17 different scripts. These materials have a limited lifespan, and as generations pass, these materials deteriorate, and the vital knowledge is lost. This work uses digital texts to convey information to future generations. Optical Character Recognition (OCR) helps extract information from scanned manuscripts (printed text). This paper proposes a simple and effective solution of optical character recognition (OCR) Sanskrit Character from text document images using long short-term memory (LSTM) and neural networks of Sanskrit Characters. Existing methods focuses only upon the single touching characters. But our main focus is to design a robust method using Bidirectional Long Short-Term Memory (BLSTM) architecture for overlapping lines, touching characters in middle and upper zone and half character which would increase the accuracy of the present OCR system for recognition of poorly maintained Sanskrit literature.

Sign in / Sign up

Export Citation Format

Share Document