A Comprehensive PhotoOCR (Scene Text Recognition) Toolkit
across 13 Indian Languages

---

IndicPhotoOCR is an advanced PhotoOCR toolkit designed for detecting, identifying, and recognizing text across 13 Indian languages, including Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Meitei Odia, Punjabi, Tamil, Telugu, Urdu, and English. Built to handle the unique scripts and complex structures of Indian languages, IndicPhotoOCR provides robust detection and recognition capabilities, making it a valuable tool for processing multilingual documents and enhancing document analysis in these diverse scripts.

---

Table of Content

Updates
Installation
How to use
Bharat Scene Text Dataset
Contributors
Acknowledgement
Contact us

---

Updates

[December 2024]: Detection Module: TextBPN++ added.
[November 2024]: Demo available in huggingface space.
[November 2024]: Code available at Google Colab.
[November 2024]: Added support for 10 languages in the recognition module.
[September 2024]: Private repository created.

---

Installation

Currently we need to manually create virtual environemnt.

conda create -n indicphotoocr python=3.9 -y
conda activate indicphotoocr


git clone https://github.com/Bhashini-IITJ/IndicPhotoOCR.git
cd IndicPhotoOCR

CPU Installation

python setup.py sdist bdist_wheel
pip install dist/IndicPhotoOCR-1.2.0-py3-none-any.whl[cpu]

CUDA 11.8 Installation

python setup.py sdist bdist_wheel
pip install ./dist/IndicPhotoOCR-1.2.0-py3-none-any.whl[cu118] --extra-index-url https://download.pytorch.org/whl/cu118

CUDA 12.1 Installation

python setup.py sdist bdist_wheel
pip install ./dist/IndicPhotoOCR-1.2.0-py3-none-any.whl[cu121] --extra-index-url https://download.pytorch.org/whl/cu121

If you find any trouble with the above installation use the setup.sh script.

chmod +x setup.sh
./setup.sh

Config

Currently this model works for hindi v/s english script identification and thereby hindi and english recognition.

Detection Model: TextBPN++
ScripIndetification Model: Hindi v/s English
Recognition Model: Hindi, English, Assamese, Bengali, Gujarati, Marathi, Odia, Punjabi, Tamil, Telugu.

How to use

Detection

>>> from IndicPhotoOCR.ocr import OCR
# Create an object of OCR
>>> ocr_system = OCR(verbose=True) # for CPU --> OCR(device="cpu")

# Get detections
>>> detections = ocr_system.detect("test_images/image_141.jpg")

# Running text detection...
# 4334 text boxes before nms
# 1.027989387512207

# Save and visualize the detection results
>>> ocr_system.visualize_detection("test_images/image_141.jpg", detections)
# Image saved at: test.png

Cropped Word Recognition

>>> from IndicPhotoOCR.ocr import OCR
# Create an object of OCR
>>> ocr_system = OCR(verbose=True) # for CPU --> OCR(device="cpu")
# Get recognitions
>>> ocr_system.recognise("test_images/cropped_image/image_141_0.jpg", "hindi")
# Recognizing text in detected area...
# 'मण्डी'

End-to-end Scene Text Recognition

>>> from IndicPhotoOCR.ocr import OCR
# Create an object of OCR
>>> ocr_system = OCR(verbose=True) # for CPU --> OCR(device="cpu")
# Complete pipeline
>>> ocr_system.ocr("test_images/image_141.jpg")
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: रोड
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: बाराखम्बा
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: राजीव
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: चौक
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: मण्डी
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Recognized word: हाऊस
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: rajiv
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: chowk
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: mandi
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: house
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: barakhamba
# Identifying script for the cropped area...
# Recognizing text in detected area...
# Using cache found in /root/.cache/torch/hub/baudm_parseq_main
# Recognized word: road

Related Large-Scale Indian Language Scene Text Dataset

Bharat Scene Text Dataset - BSTD

Contributors

Anik De
Tech Lead and Main Contributor

Abhirama	Aditya Rathore	Harshiv Shah
Contributor	Contributor	Contributor

Sagar Agarwal	Rajeev Yadav
Contributor	Contributor

Anand Mishra
Project Investigator

Citation

@misc{ipo,
  author = {Anik De et al.}
  title      = {{I}ndic{P}hoto{O}CR: A comprehensive toolkit for {I}ndian language scene text understanding},
  howpublished = {\url{https://github.com/Bhashini-IITJ/IndicPhotoOCR/}},
  year         = 2024,
}

Acknowledgement

Text Recognition - PARseq
EAST re-implemenation repository.
TextBPN++ repository.
National Language Translation Mission Bhashini.

Contact us

For any queries, please contact us at:

• Anik De