Sinhala Printed OCR - Custom Tesseract Model

This repository contains a custom-trained Tesseract OCR model for recognizing printed Sinhala text.

The main trained model file is:

sin_custom_v3.traineddata

This model was trained and evaluated as part of a Sinhala OCR research component for extracting Sinhala text from educational documents.

---

Repository Contents

.
├── sin_custom_v3.traineddata     # Custom trained Tesseract Sinhala OCR model
├── run-wer-cer.ipynb             # Notebook used to evaluate WER and CER
└── README.md                     # Project documentation

---

Model Description

sin_custom_v3.traineddata is a custom Tesseract LSTM OCR model trained for printed Sinhala text recognition.

The model was trained by the researcher using a Sinhala OCR training dataset containing both printed and handwritten Sinhala samples. For this Tesseract model, the evaluation was focused only on the printed Sinhala samples.

---

Dataset

The dataset used for training and evaluation was:

Sinhala OCR Training Data
Kaggle Dataset:
https://www.kaggle.com/datasets/hasindukoshitha/sinhala-ocr-training-data

Dataset summary:

Dataset Detail	Value
Printed samples	2,840
Successfully processed printed samples	2,829

---

Training Summary

The custom Tesseract model was trained using the Tesseract LSTM training pipeline. During training, the model showed a significant reduction in internal training error rates.

Metric	Value
Model name	sin_custom_v3
OCR engine	Tesseract LSTM
Output classes	117
Total model weights	940,693
Initial BCER	22.239%
Initial BWER	50.684%
Best BCER	4.167%
BWER at best checkpoint	14.322%

The best model checkpoint was selected based on the minimum training Best Character Error Rate (BCER).

---

Evaluation Results

The model was evaluated using the notebook:

run-wer-cer.ipynb

The notebook loads the trained Tesseract model, runs OCR on printed Sinhala image samples, and compares the OCR output with the ground-truth text from metadata.csv.

Evaluation was performed using:

• WER - Word Error Rate
• CER - Character Error Rate

Final printed OCR evaluation results:

Metric	Value
Average WER	0.1574
Average WER percentage	15.74%
Average CER	0.0346
Average CER percentage	3.46%

---

Result Interpretation

The trained Tesseract model achieved a Character Error Rate of 3.46% and a Word Error Rate of 15.74% on printed Sinhala text samples.

The low CER indicates that the model performs well at recognizing individual Sinhala characters. However, the WER is higher because a word is counted as incorrect even when only one character inside the word is wrongly recognized.

This is especially important in Sinhala OCR because small character, vowel sign, or diacritic errors can change the full word.

---

How to Use the Model

1. Copy the traineddata file

Copy sin_custom_v3.traineddata into your Tesseract tessdata directory.

Example on Windows:

C:\Program Files\Tesseract-OCR\tessdata

Example on Linux:

/usr/share/tesseract-ocr/4.00/tessdata/

or:

/usr/share/tesseract-ocr/5/tessdata/

---

2. Run Tesseract from command line

tesseract input_image.png output -l sin_custom_v3 --oem 3 --psm 6

This will generate:

output.txt

---

3. Use with Python

import pytesseract
from PIL import Image

image = Image.open("sample_image.png")

text = pytesseract.image_to_string(
    image,
    lang="sin_custom_v3",
    config="--oem 3 --psm 6"
)

print(text)

---

Evaluation Notebook

The run-wer-cer.ipynb notebook performs the following steps:

1. Loads the dataset metadata.
2. Filters printed Sinhala samples.
3. Runs OCR using sin_custom_v3.
4. Compares predicted text with ground-truth text.
5. Calculates WER and CER using the jiwer library.
6. Reports average WER and CER.

---

Requirements for Evaluation

Install the required Python libraries:

pip install pytesseract pillow pandas jiwer matplotlib

Tesseract OCR must also be installed on the system.

---

Limitations

This model is mainly designed for printed Sinhala OCR.

Current limitations include:

• Not optimized for handwritten Sinhala text.
• May struggle with noisy, blurred, skewed, or low-resolution images.
• May produce errors for unsupported symbols, mixed English-Sinhala text, and punctuation.
• Word-level accuracy is lower than character-level accuracy.
• Further improvement is needed for full-page OCR with complex layouts.

---

Future Improvements

Future improvements may include:

• Expanding the Sinhala character set.
• Cleaning unsupported symbols from training text.
• Adding more printed Sinhala font variations.
• Improving preprocessing for noisy document images.
• Evaluating with a separate unseen test dataset.
• Combining Tesseract printed OCR with a separate handwritten OCR model.

---

Research Use

This model was developed for Sinhala OCR research and educational document processing. It can be used in Sinhala document digitization pipelines, educational assistant systems, and OCR-based search or question-answering systems.

---

Model Version

Model name: sin_custom_v3
File: sin_custom_v3.traineddata
OCR engine: Tesseract LSTM
Task: Printed Sinhala OCR
Evaluation CER: 3.46%
Evaluation WER: 15.74%

---

Citation

If this model or dataset is used in academic work, please cite the dataset and mention that the custom Tesseract OCR model was trained for printed Sinhala text recognition.

Dataset:

H. Koshitha, "Sinhala OCR Training Data," Kaggle Dataset.
Available: https://www.kaggle.com/datasets/hasindukoshitha/sinhala-ocr-training-data