# Bangla OCR
- An under explored field, no good deep learning based model available.

- Best online API I know is the Google OCR API, which is very robust and accurate.

- Huge potential in such a model, since many companies need to digitize scanned documents and also mobile apps with Bangla OCR will also have great usage for the general public.

## Goal
To develop a robust Bangla OCR model with the least amount of labelled data.

## Scope of the work

1. Classification in Bangla
2. Summarization in Bangla
3. Name Entity Recognition in Bangla
- Detection of an entity
- Categorization of an entity
- Entity means person, organization, and/or lcoation
- For eg. in a sentence, we write আমি লালমনিরহাটে লালমনির স্টোরে গিয়ে লালমনিরের সাথে দেখা করলাম।
- For more info see: https://turkunlp.org/fin-ner.html This is only a finnish demonstration. English have more contexts in this regard.
5. Understanding documents (2D structure) Chargrid representation for eg.

## Existing works
- A review paper on Bangla language: https://arxiv.org/abs/2105.14875 submitted 2021.
- There are many OCR apps available if you simply google "Bangla OCR".
- But almost none are DL based, and there are some research papers which are trained on little data.
- Excellent link for understanding NLP models: https://jalammar.github.io/
- Excellent link for transformer: https://jalammar.github.io/
- Reference/ review paper on: https://arxiv.org/abs/2012.12556?fbclid=IwAR2f7lkdKsvf-LqLTYDB65kezkfBXuUxGqcP31jo1dq1kO6fql4fdN0hCT0
- Bangla Glove: https://github.com/sagorbrur/GloVe-Bengali

## Information on various CapsNet architecture and implementation:
https://github.com/loretoparisi/CapsNet/blob/master/README.md

## Bangla Existing corpous links
- Corpous data https://github.com/banglakit/awesome-bangla?fbclid=IwAR1LFME30Aw418f_icnNp-SJVGtmRNV7Jii8mGnQBsv_etv_fqdlQaKwTr4#corpora-corpus-and-datasets
- Bangla wiki https://dumps.wikimedia.org/bnwiki/latest/?fbclid=IwAR2cPIoxDPQUHeiPP3HgWZDjJcOnjH75XvhN03t6CiMlc1JhVdL3AXMY-28

### @QH2yNM7OQ8OoWgMjSLjIXQ Can you list those app here?
- app1 (link)
- app2 (link)

## What we know about Optical Character Recognition(OCR)?
- What is needed to make an OCR?
- Template matching based OCR

## Good OCR models from English
- English OCR has the best available models, example `keras-ocr`: https://github.com/faustomorales/keras-ocr
- It is deep learning based, and also comes with a custom training script: https://keras-ocr.readthedocs.io/en/latest/examples/end_to_end_training.html
- Tessaract (LSTM) custom training: https://tesseract-ocr.github.io/tessdoc/TrainingTesseract-4.00.html
- Apache tika: https://medium.com/@masreis/text-extraction-and-ocr-with-apache-tika-302464895e5f

## Handwritten OCR is too hard
- Human handwriting is too variable and noisy.
- A simpler solution exists for printed letters, since synthetic data can easily be generated for the most common fonts.
- Handwriting can't be synthetically generated like Bangla fonts.

## Ideas for developing Bangla OCR
1. Collect many popular Bangla fonts online.
2. Create a synthetic image data generator that creates images and labels for the bangla fonts, example generator: https://github.com/Belval/TextRecognitionDataGenerator
3. Augment the data with things like colors, brightness, noise and picture backgrounds.
4. Train a model on huge amounts of synthetic data until robust and practical.
5. Further train and fine tune on real world datasets (e.g. scanned documents, streetview signs).
6. Even further training using an active learning system.
7. Feed the OCR predictions into the chargrid network for document classification

## Related Ideas relevant to this project
1. Category of list of shops
2. List of shops in the city/BD/area . If this solved, 1. Can be solved..
3. Building a data project for housing market. (See www.etuovi.fi search for a place name 'klaukkala-nurmijarvi' you can see list of houses for sell or rent.)
4. Later expand this project to automate the service to do (docusign style) authorization.
5. Segment generator https://github.com/farhanhubble/udacity-connect/blob/master/segmented-generator.ipynb and https://medium.com/100-shades-of-machine-learning/https-medium-com-100-shades-of-machine-learning-rediscovering-semantic-segmentation-part1-83e1462e0805

## Chargrid representation requirements
Bounding box position information can be extracted from
- OCR
- directly from layout of pdf/html

## References
1. A survey on optical character recognition for Bangla and Devanagari scripts. https://www.researchgate.net/publication/257768180_A_survey_on_optical_character_recognition_for_Bangla_and_Devanagari_scripts