ML Tasks Image OCR Calamari

[junxnone]

Reference

• paper - 2018 Calamari A High-Performance Tensorflow-based Deep Learning Package for Optical Character Recognition
• paper - 2018 - State of the Art Optical Character Recognition of 19th Century Fraktur Scripts using Open Source Engines
• paper - 2017 - Improving OCR Accuracy on Early Printed Books by utilizing Cross Fold Training and Voting
• Github - Calamari-OCR/calamari
• 字符识别引擎Calamari总结
• ocreval
• The ISRI analytic tools for OCR evaluation.pdf
• OCR-Workshop - CIS OCR Group
• Calamari-OCR Voting

Brief

Calamari-OCR Voting

• code analysis
• character level
• textline
  • voting length

---

• Single - 取单个 Confidence 最大值 频率最大的
• Adding - 取多个 Confidence 和最大的

• 每行代表一个 voter 对字符 x predict 为 [l/Ⅰ/L] 的 Confidence
  • Single - 有两个 voter predict 为 Ⅰ
  • Adding - 三个 voter 相同字符的 Confidence (列)相加后 l Confidence 最大

Model Cross Fold Training

• data 分为 N 份
• 1-N 每次选择一份作为 Testing，其他作为 Training，训练 N 个 Model

  N = 5

Code Analysis

• sequence_voter
• confidence_voter_default_ctc
• confidence_voter_fuzzy_ctc

  fuzzy_ctc 没用到 是什么？

[junxnone]

junxnone/aiwiki#200