A classifier to determine page quality from an Optical Character Recognition (OCR) perspective is developed. It classifies a given page image as either ``good'' (i.e., high OCR accuracy is expected) or ``bad'' (i.e., low OCR accuracy expected). The classifier is based upon measuring the amount of white speckle, the amount of broken pieces, and the overall size information in the page. Two different sets of test data were used to evaluate the classifier: the Sample 2 dataset containing 439 pages and the Magazines dataset containing 200 pages. The classifier recognized 85%of the pages in the Sample 2 correctly. However, approximately 40%of the low quality pages were misclassified as ``good.'' To solve this problem, the classifier was modified to reject pages containing tables or less than 200 connected components. The modified classifier rejected 41%of the pages, correctly recognized 86%of the remaining pages, and did not misclassify any low quality page as ``good''. Similarly, it recognized 86.5%of the pages in the Magazine dataset correctly and did not misclassify any low quality page as ``good'' without any rejections.

Downloadable versions:

• Postscript version (4.7MB) ms_thesis.ps
• Gzipped Postscript version (530K) ms_thesis.ps.gz

---

Contents

• Acknowledgements
• Contents
• List of Tables
• List of Figures
• Introduction
  • Optical Character Recognition
  • Needs for Estimating Page Quality
  • Difficulty in Predicting OCR Accuracy
  • Problem Description
    • Assumptions
  • Description of the Work
• Related Work
  • OCR Difficulty Evaluation
  • Other Approaches
• Classifier Design
  • Justification for Using Simple Features
  • Feature Selection Process
    • Concept Exploration Dataset
      • Concept Exploration Observations
      • Connected Components
    • Feature Metrics Design
      • White Speckle
      • Broken-Chars Zone
      • Size Information
    • Preliminary Set of Rules
  • Determining Threshold Values
    • Training Dataset
    • Conclusions from Training Test
      • White Speckle
      • Reference Point
      • Broken Chars Zone
      • Size Information
  • Final Set of Rules
  • Summary
• Results and Analysis
  • Classifier Testing Architecture
    • Basic Processing Model
  • Test Data Set
  • Unfiltered Results
  • Error Analysis
  • Higher Good Thresholds
    • Good Threshold = 95%Results
    • Good Threshold = 98%Results
  • Magazine Data
  • Summary
• Future Work
  • Statistical Pattern Recognition
  • Features Observed but not Used
  • New Datasets
• Concluding Remarks
• Classifier Results for Test Dataset
• Classifier Results for Magazine Dataset
• References
• About this document ...