https://github.com/cjbarrie/noisy-ocr-benchmark

Replication materials for the article "OCR with Tesseract, Amazon Textract, and Google Document AI: A Benchmarking Experiment"

https://github.com/cjbarrie/noisy-ocr-benchmark

Science Score: 10.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • CITATION.cff file
  • codemeta.json file
  • .zenodo.json file
  • DOI references
  • Academic publication links
    Links to: zenodo.org
  • Academic email domains
  • Institutional organization owner
  • JOSS paper metadata
  • Scientific vocabulary similarity
    Low similarity (5.3%) to scientific vocabulary
Last synced: 10 months ago · JSON representation

Repository

Replication materials for the article "OCR with Tesseract, Amazon Textract, and Google Document AI: A Benchmarking Experiment"

Basic Info
  • Host: GitHub
  • Owner: cjbarrie
  • Default Branch: main
  • Size: 60.1 MB
Statistics
  • Stars: 0
  • Watchers: 0
  • Forks: 0
  • Open Issues: 0
  • Releases: 0
Fork of Hegghammer/noisy-ocr-benchmark
Created almost 5 years ago · Last pushed about 5 years ago

https://github.com/cjbarrie/noisy-ocr-benchmark/blob/main/

# An OCR Benchmarking Experiment
  
This repository holds replication materials for the manuscript ["OCR with Tesseract, Amazon Textract, and Google Document AI: A Benchmarking Experiment"](https://osf.io/preprints/socarxiv/6zfvs). It contains:

- The .RMD file of the manuscript with R code for all the figures.
- 51,304 .TXT files with the text output from all the OCR processing requests.
- A .CSV file with word and character accuracy rates for all the OCR output.

The image test materials reside in a separate Zenodo repository as the ["Noisy OCR Dataset"](https://zenodo.org/record/5068735) (NOD).

From the abstract: 

>This article reports a benchmarking experiment comparing the performance of Tesseract, Amazon Textract, and Google Document AI on images of English and Arabic text. English-language book scans (n=322) and Arabic-language article scans (n=100) were replicated 43 times with different types of artificial noise for a corpus of 18,568 documents, generating 51,304 process requests. Document AI delivered the best results, and the server-based processors (Textract and Document AI) were substantially more accurate than Tesseract, especially on noisy documents. Accuracy for English was considerably better than for Arabic.

Core results: 
\
\
![](images/fig4.png)

Owner

  • Name: Christopher Barrie
  • Login: cjbarrie
  • Kind: user
  • Company: University of Edinburgh

Lecturer in Computational Sociology, University of Edinburgh.

GitHub Events

Total
Last Year