myOCR (Optical Character Recognition Corpus for Myanmar Language or Burmese)

myOCR is a comprehensive synthetic dataset for Optical Character Recognition (OCR) of the Myanmar (Burmese) language. This resource addresses the critical shortage of open, high-quality datasets for Myanmar OCR, a significant barrier to developing robust document digitization tools for this low-resource language.

We generated 25,790 synthetic text-line images using 14 diverse Myanmar font styles to ensure broad coverage and reduce data bias. The dataset supports end-to-end OCR model training and evaluation, from optical character recognition to post-processing error correction. Each image is paired with a corresponding word-segmented text label.

Experimental Findings

• The best-performing optical model (9,000 iterations) achieved a CHRF++ score of 97.90% and a Word Error Rate (WER) of 9.18%.
• Post-OCR correction significantly improved results:
• Transformer-based correction reduced WER to 0.66% and increased CHRF++ to 99.31%.
• Statistical methods (N-gram/SymSpell) also reduced WER by nearly fivefold when combined with sequence modeling.
• Error analysis revealed unique challenges in Myanmar OCR, including visual rendering vs. logical character order issues.

Data Format

The Burmese text data are word-segmented using underscore ("_") as the delimiter. The dataset follows this format:

<img-file-path>\t<text-label>

Example from train.txt:

ye@lst-hpc3090:~/tmp/myOCR/data/ver1.0$ head train.txt
Images/MasterpieceUniType_312.png       တပို့တွဲ_လ_တွင်_၂၉_ရက်_ရှိ_သည်
Images/Press_1127.png   အားနာ_စရာ_ကြီး_အကြာကြီး_ရှာ_နေ_ရ_တယ်
Images/MyanmarAyar3_664.png     စိတ်မပူ_ပါ_နဲ့_ရှာ_တွေ့_အောင်_ကြိုးစား_ပါ_မယ်
Images/UMoe_1231.png    ဘူတာ_ကြီး_အချို့_မှာ_သာ_ရထား_ကို_ရေ_ဖြည့်_လို့_ရ_တာ
Images/NKSS_164.png     သူ_ပျင်း_လာ_ရင်_အချိန်_မ_ရှိ_အခါ_မ_ရှိ_ဖုန်းဆက်_တယ်
Images/B_HW_830.png     ဈေးနှုန်း_က_ဘယ်လောက်_လဲ
Images/MyanmarYinMar_1344.png   ကိစ္စ_မ_ရှိ_ပါ_ဘူး
Images/MyanmarYinMar_42.png     ကျောင်း_မှာ_မှတ်ပုံတင်ကြေး_ပေး_သည်
Images/MasterpieceUniType_1496.png      ရှန်ဟိုင်း_မြို့_ရဲ့_လူဦးရေ_ဟာ_ဘယ်လောက်_လဲ
Images/B_HW_440.png     အမေ_က_ခင်ဗျား_ကို_သူ_နဲ့အတူ_နေ_ခွင့်_မ_ပေး_ဘူး
ye@lst-hpc3090:~/tmp/myOCR/data/ver1.0$

Example from valid.txt:

ye@lst-hpc3090:~/tmp/myOCR/data/ver1.0$ head valid.txt
Images/MasterpieceUniType_312.png       တပို့တွဲ_လ_တွင်_၂၉_ရက်_ရှိ_သည်
Images/Pyidaungsu_207.png       နောက်_အပတ်_စနေ_နေ့_မှာ
Images/Press_1365.png   နောက်ဆုံး_ရထား_က_ဘယ်_အချိန်_ရှိ_ပါ_သလဲ
Images/B_HW_2295.png    နောက်_တစ်_ခေါက်_ထပ်_ဆို_ပါ_ဦး
Images/MasterpieceUniType_2125.png      ကိုင်_ထား_ပါ
Images/UMoe_1890.png    ခင်ဗျား_ဘယ်_နေရာ_ကို_ကြည့်_ချင်_လဲ
Images/Press_287.png    အာရုဏ်ခင်း_လေ_က_တကယ်_ကို_ပဲ_သန့်ရှင်း_လတ်ဆတ်_သည်
Images/MasterpieceUniType_838.png       ကျွန်တော်_က_ဆိုးလ်_မှာ_နေ_ပါ_တယ်
Images/MyanmarSanpya_627.png    ဟွန်း_က_အလုပ်_မ_လုပ်_ဘူး
Images/Pyidaungsu_639.png       မကြာခဏ_တွေ့ဆုံ_ရင်_အနေနီး_လာ_တာ_သဘာဝ_ပဲ
ye@lst-hpc3090:~/tmp/myOCR/data/ver1.0$

Archive with 7z

The line-level text images for OCR are stored in the data folder and compressed using 7z. Due to GitHub's file size limitations, we used the -v52m option to split the archive into 52MB volumes.

ye@lst-hpc3090:~/tmp$ time 7z a -v52m Images.7z ./Images/

7-Zip 23.01 (x64) : Copyright (c) 1999-2023 Igor Pavlov : 2023-06-20
 64-bit locale=en_US.UTF-8 Threads:24 OPEN_MAX:1024

Scanning the drive:
1 folder, 25790 files, 91981953 bytes (88 MiB)

Creating archive: Images.7z

Add new data to archive: 1 folder, 25790 files, 91981953 bytes (88 MiB)


Files read from disk: 25790
Archive size: 85800971 bytes (82 MiB)
Volumes: 2
Everything is Ok

real    0m4.250s
user    0m8.873s
sys     0m0.260s
ye@lst-hpc3090:~/tmp$

The compression resulted in two split archive files:

ye@lst-hpc3090:~/tmp$ ls -lh Images.7z.*
-rw-rw-r-- 1 ye ye 52M Aug 19 20:13 Images.7z.001
-rw-rw-r-- 1 ye ye 30M Aug 19 20:13 Images.7z.002
ye@lst-hpc3090:~/tmp$

To extract all images, simply unzip the first volume file. The 7z utility will automatically handle the multi-volume extraction:

7z x Images.7z.001

Text-to-Image Conversion Script

We used the mytext2pic.sh script to convert Myanmar sentences into PNG image files. The code is as follows:

#!/bin/bash

# make "png files" for each Myanmar sentence from the input file
# written by Ye Kyaw Thu, Language Understanding Lab., Pyin Oo Lwin, Myanmar
# Note: you need "mytext.tex" latex file and your file that contained Myanmar sentences
# under the same folder with "mytext2pic.sh"
#
# Last updated: 19 June 2020 
# How to run: bash ./mytext2pic.sh <input-file>
# e.g. bash ./mytext2pic.sh ./mytxt4png.txt
# output files: line1.png, line2.png, line3.png ...

# Ref: https://tex.stackexchange.com/questions/34054/tex-to-image-over-command-line/34058#34058
# Ref: http://phyletica.org/tex-equations-on-cmd-line/

count=1;

cat $1 | while read -r line
do
   xelatex "\def\mytext{${line}}\input{mytext.tex}"
   convert -density 300 mytext.pdf -quality 90 line$count.png
   ((count++))
done

Random Samples

We used 14 different Myanmar fonts for our experiments. Two random text image samples for each font are shown in the table below:

Table 1: Two randomly selected text image samples for each of the 14 Myanmar fonts utilized in the study

No	Font name	Text Image Sample-1	Text Image Sample-2
1	Burmese Handwriting Style 04
2	Myanmar Chatulight
3	Kamjing
4	Z09-LatYaySat
5	Masterpiece Spring Revolution
6	Masterpiece Uni Type
7	Myanmar Ayar3
8	Myanmar Phiskel
9	Myanmar Sanpya
10	Myanmar Yin Mar
11	NKSSmart3
12	Z03-Press
13	Pyidaungsu
14	Z01-UMoe

Distribution of Font Usage in Synthetic Text Images

The following figure illustrates the distribution of text lines across different fonts used in the synthetic dataset.

Fig.1 Number of lines of each font used in the synthetic text images

Hugging Face Link

We first uploaded the myOCR version 1.0 dataset to Hugging Face on November 18, 2024. You can access the myOCR dataset at the following link:
https://huggingface.co/datasets/LULab/myOCR

Conference Presentation

Our work on POST-OCR error correction with the myOCR dataset was presented at the iSAI-NLP 2024 conference. The presentation slides can be accessed at: https://github.com/ye-kyaw-thu/myOCR/blob/main/doc/slide/myOCR_isai21_v2.pdf

Citation

Please cite this if you use this dataset:

Thura Aung, Ye Kyaw Thu and Myat Noe Oo, "myOCR: Optical Character Recognition for Myanmar language with Post-OCR Error Correction", In Proceedings of the 19th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP 2024), Nov 11 to 15, 2024, Pattaya, Thailand.

Abstract:
This paper presents the Myanmar Optical Character Recognition (OCR), named myOCR. It utilizes a synthetic text image dataset with 14 different font styles that contains 25,790 text images. The system includes Convolutional Neural Networks (CNN) for feature extraction, Bidirectional Long-Short Term Memory (BiLSTM) networks for sequence modeling, and Connectionist Temporal Classification (CTC) for decoding, evaluated across various iterations (3,000, 6,000, 9,000) and hidden states (64, 128, 256). Statistical Post-OCR correction methods involve N(3,4,5)-grams and edit distances with the Symmetric Delete Spelling correction algorithm (SymSpell). For Neural Machine Translation-based correction, BiLSTM and Transformer models are employed, while the mT5-base and mBART-50 models are used for LLM-based correction. The best base (optical) model is the model with 9,000 iterations that achieved a chrF++ score of over 97.90 and a Word Error Rate (WER) of 9.18%. Transformer correction improved its chrF++ to 99.31 and reduced the WER to 0.66%.

Paper link: https://ieeexplore.ieee.org/document/10799448