Thai National Document Optical Character Recognition (THND OCR)

Tesseract OCR tools for read Thai National Document used TH Sarabun National Font trained and fine-tuned. Read README.md to see about my process.

• Part I : https://github.com/copninich/TH-National-Document-OCR-Part-I

• Part II : https://github.com/copninich/TH-National-Document-OCR-Part-II

• Medium : https://medium.com/@copninich/5a673cc8a686

0. Information

0.1 Tool

• Tesseract : https://github.com/tesseract-ocr/tesseract

0.2 Datasets

• PyThaiNLP (Prachathai) : https://github.com/PyThaiNLP/prachathai-67k
• PyThaiNLP (ThaiGov V2 Corpus) : https://github.com/PyThaiNLP/thaigov-v2-corpus
• PyThaiNLP (ThaiGov Archive corpus) : https://github.com/PyThaiNLP/thaigov-archive-corpus
• Thaisum : https://github.com/nakhunchumpolsathien/ThaiSum
• TR-TPBS : https://github.com/nakhunchumpolsathien/TR-TPBS

0.3 Project information

• Ai Builders 2021
• Kampanart Chaimooltan

01.Performance tested

I used Character Errorate and leght string (OCR & Correct Text) and output result testing (.csv file)

02.Generated datasets

I used PIL library. in addtion, I used TH Sarabun formart font 72 px to create datasets.

Link : https://www.kaggle.com/copninich/thaienglish-character-in-th-sarabun-font

03.Tested trained and fine-tuned Tesseract (default langdata_lstm)

Requirements

1. langdata_lstm
2. tesseract v.4
3. tessdata_best

Load file to your folder and extract : https://drive.google.com/drive/folders/1ABo7ooO62Tb03RR_VvkdshRVG9vz23sl?usp=sharing

04.Trained and fine-tuned

Run script script_basic.ipynb or script_config_error.ipynb

Requirements

1. langdata_lstm
2. tesseract v.4
3. tessdata_best

Custom tha.training_text with my own datasets more than 1.9 M sentences

05.Performance tested

report_performace_final.csv