Lehigh University - CSE: Baird 2

Dr. Henry Baird

"Automatic Recognition of Images of Extremely Poor Quality"

Thursday, September 23, 4 PM

Packard Lab, Room 466

Abstract: Low image quality is a serious obstacle to highly accurate automatic recognition of text. We show that document image decoding (DID) algorithms, as a result of recent refinements, perform well despite severe degradation in both training and test images. This is due to refined algorithms for DID training of character-template, set-width, and channel (noise) models. Large-scale experiments using synthetically degraded images have shown that:
1) high accuracy (>99% characters correct) is achievable using DID models
   trained on severely degraded images from the same distribution; and
2) greatly improved accuracy (<1/10 the error rate) results across a wide range
    of image degradations.
We also propose a recognition methodology using 'decoder banks' that accomplishes all of the above without requiring manual intervention or document-specific supervised training.   This is possible since DID decoders
1) are trainable for high accuracy across a wide range of explicitly parameterized
   image degradations, and
2) can be generated fully automatically for arbitrary parameter settings.
When implemented naively in a brute-force manner, decoder banks are computationally
intensive: but we suggest ways that this cost may be reduced with no loss of versatility,
automation, or accuracy.

(This work is joint with Prateek Sarkar of PARC. This combines two talks originally given at ICDAR 2003 (Edinburgh) and ICPR 2004 (Cambridge, UK).)

Bio: Dr. Baird is Professor of Computer Science & Engineering at Lehigh Univ. and (with Dan Lopresti) heads up Lehigh's Pattern Recognition Research lab. He is a Fellow of the IEEE and also of the IAPR, and received the 2003 ICDAR Outstanding Contributions award. He's a founding member of the Editorial Board of the Int'l J. on Document Analysis and Recognition. He has published three books and over seventy technical articles, and holds
seven patents.

		Dr. Henry Baird "Automatic Recognition of Images of Extremely Poor Quality" Thursday, September 23, 4 PM Packard Lab, Room 466 Abstract: Low image quality is a serious obstacle to highly accurate automatic recognition of text. We show that document image decoding (DID) algorithms, as a result of recent refinements, perform well despite severe degradation in both training and test images. This is due to refined algorithms for DID training of character-template, set-width, and channel (noise) models. Large-scale experiments using synthetically degraded images have shown that: 1) high accuracy (>99% characters correct) is achievable using DID models trained on severely degraded images from the same distribution; and 2) greatly improved accuracy (<1/10 the error rate) results across a wide range of image degradations. We also propose a recognition methodology using 'decoder banks' that accomplishes all of the above without requiring manual intervention or document-specific supervised training. This is possible since DID decoders 1) are trainable for high accuracy across a wide range of explicitly parameterized image degradations, and 2) can be generated fully automatically for arbitrary parameter settings. When implemented naively in a brute-force manner, decoder banks are computationally intensive: but we suggest ways that this cost may be reduced with no loss of versatility, automation, or accuracy. (This work is joint with Prateek Sarkar of PARC. This combines two talks originally given at ICDAR 2003 (Edinburgh) and ICPR 2004 (Cambridge, UK).) Bio: Dr. Baird is Professor of Computer Science & Engineering at Lehigh Univ. and (with Dan Lopresti) heads up Lehigh's Pattern Recognition Research lab. He is a Fellow of the IEEE and also of the IAPR, and received the 2003 ICDAR Outstanding Contributions award. He's a founding member of the Editorial Board of the Int'l J. on Document Analysis and Recognition. He has published three books and over seventy technical articles, and holds seven patents.