[status]
The goal is to build a next generation of OCR technology. Eventually, it helps to target for a large-scale, open source, global language and culture data bank.
This package contains an OCR engine specifically for Pashto, Persian, and Traditional Chinese. Currently it consists of major components: segmentation, layout analysis and OCR. Its OCR function uses certain parts of latest Tesseract OCR (https:\//github.com/tesseract-ocr/tesseract) for character recognition.
Many implementations of OCR software exist, both commercial and open source, but they do not produce useful results for Traditional Chinese, Pashto, and Persian literature. 90%+ accuracy rate is a threshold for OCR to be useful. Some of the most spectacular failures of OCR software are not a result of its inability to recognize characters, but are unable to perform accurate layout analysis by identifying text regions, images, and text direction.
The proposed advanced implementation will be a next generation of OCR software capable of handling complex layouts.
The project is currently funded by the National Endowment of Humanities (NEH) for $75,000 for the period of 2019.
Marek Rychlik is the Principal Investigator. Yan Han and Dylan Murphy are the Co-PI. Other project staff include Raymundo Navarette, Dwight Nwaigwe, Sayyed Mohsen Vazirzade.
The current team started studying the OCR problem in Fall 2017. Marek approached Yan about the Afghan Digital Collection, the largest digital repository related to Afghanistan and related areas, at the University of Arizona Libraries. Yan Han and his colleague Atifa Rawan have been working with ACKU for the past 11 years on providing access and preservation of Afghanistan materials. To date, the Afghanistan Digital Collection provides preservation of and access to an extensive collection of Afghan related materials (1.7+ million pages) and 500k pageviews in usage since its first availability in 2010.
In 2017, the project team started a regular seminar devoted to OCR with emphasis on these languages. We surveyed available OCR systems and most successful algorithms. It turns out that the best algorithm of today is based on Long-Short Memory (LSTM) recurrent neural network (RNN), and which is implemented in the two most accurate OCR systems when applied to Arabic: Kraken and Tesseract. We evaluated both systems for accuracy and studied the methodology of extending these systems to handle Pashto.
We set up a Subversion repository of test data, code written in MATLAB, and a collection of relevant research papers. We investigated the theory of LSTM and discovered certain new RNNs which may be better for OCR than LSTM (to be further studied). We also added Traditional Chinese to our studies, a different writing system, in order to increase our understanding of common and distinct features of various writing systems.
Our preliminary research suggests that current OCR has low accuracy especially on Traditional Chinese texts printed before the 1900s.We developed a new algorithm which in tests shows the ability to recognize several common Chinese characters in a 118 page document with very high accuracy (> 95%). A video used in the visualization of the search for a particular character ( ).
Furthermore, we discovered that current OCR systems trip up on the initial phase of processing, image segmentation, whose goal is to isolate characters by subdividing areas of the image occupied by different characters into regions.
2017 - , OCR seminars are held every Friday. Information is available at (http://alamos.math.arizona.edu/ocr/)
2018, The research project was funded by the National Endowment for the Humanities (NEH) for the amount of $75,000. This Phase I we are focusing on creating prototype software for extracting complete and accurate full-text for Pashto and traditional Chinese languages. We have made good progress and would like to demonstrate the current research to you.
2019, A new NEH grant was submitted for Phase II of this project. It will focus on scalability and new language (e.g. Persian) with the use of Natural Language Processing along with Machine Learning and Artificial Intelligence.
NOTE: This package depends on other packages that may be licensed under different open source licenses.
For the latest online version of the README.md see:
Traditional Chinese: Chinese is spoken in China, Republic of China, Taiwan, Singapore, and people whose an. It is estimated that 1.2 billion people (16% of the world’s population) speak Chinese. The entire Chinese character corpus comprises 55,000 characters. Approximately 3,000 characters are required to be literacy. Each Chinese character represents a monosyllabic Chinese word or morpheme. Chinese characters have two writing systems: simplified and traditional Chinese. They are look alike, because simplified chinese took place in 1954 for the purpose of promoting mass literacy to simplify complex traditional glyphs to fewer strokes.
Persian Language: A language spoken in Iran and Afghanistan. Dari and Farsi is a variety of the Persian language spoken in Afghanistan. The modern Persian script is directly derived from Arabic script. There are 32 letters of the modern Persian alphabet along with diacritics. The appearance of a letter changes depending on its position in a word: initial, medial, final and alone. The rules can be found at the ALA-LC Romanization Tables at the Library of Congress (https://www.loc.gov/catdir/cpso/roman.html)
Pashto Language: Dari and Pashto are the two official languages of Afghanistan. Pashto is also spoken in regions of Pakistan. The total number of Pashto speakers is estimated 45 - 60 million people. Pashto alphabet is derived from Arabic script. There are 45 letters of the alphabet and 4 diacritic marks. The appearance of a letter changes depending on its position in a word: initial, medial, final and alone. The rules can be found at the ALA-LC Romanization Tables at the Library of Congress (https://www.loc.gov/catdir/cpso/roman.html)
Multiple algorithms are in consideration or in use for this project. Due to the nature of different languages, different algorithms are used. A different approach has been taken, compared to tesseract. Algorithms include: a) convert image to binary; b) compute outline cycle for each blob; c) cubic spline interpolation; d) sample every cycle with a fixed number of points; e) cross-corelation to align cycles; f) calculate the distance between aligned outlines; g) dynamic time warping; f) Recurrent Neural Networks (RNN); g) repair missing parts etc.
Chinese OCR matching using segmentation box
Matching English characters using outlines
Matching Chinese characters using outlines
Mapping one Font to another
