My Works at Apurba

My Works at Apurba Tech

Apurba technologies Inc is a USA-based software development company, with regional offices in Bangladesh and Malaysia, that creates enterprise big data solutions using NLP, machine learning, and Computer Vision.

I joined Apurba as a Machine Learning Engineer (Computer Vision) and was assigned to Bangla Optical Character Recognition Project which was delegated by Bangladesh ICT Ministry. The goal of this project was a build a complete Bangla OCR Engine deployable online with integrated handwriting recognition system.

During my time at Apurba, I was mainly involved with -

Research and development of large-scale Bengali Optical Character Recognition (OCR) pipelines.
Developed and optimized deep learning models for grapheme-based bengali character recognition.
Implemented image processing & deep learning methods for word detection & character segmentation.

Publication

[1] AKM Shahariar Azad Rabby, Md. Majedul Islam, Zahidul Islam, Nazmul Hasan, Fuad Rahman, “Towards Building A Robust Large-Scale Bangla Text Recognition Solution Using A Unique Multiple-Domain Character-Based Document Recognition Approach”, 20th IEEE International Conference on Machine Learning and Applications (ICMLA), December 13-16, 2021, (Accepted for Publication)

Word Segmentation for Bangla Document Images

Improved word detection in the OCR pipeline by replacing image processing algorithm with a deep learning based solution leveraging EAST neural network.
Modified EAST with custom CNN such as Mobilenets and fine-tuned it for Bangla Printed and Hand-written text documents.
The model takes a document image as input and outputs bounding boxes for each word in the document.

Grapheme-based Optical Character Recognition

As bangla has a complex writing system with frequent usage of vowel and consonant, we developed grapheme based deep models which predicts the three components of a bangla character - grapheme root, vowel diacritic, consonant diacritic.
Designed a mobilenet based model which is computationally inexpensive but still retains good performance indicating its potential usefulness in mobile applications.

Character Segmentation from Cropped Word Images

Developed image processing based algorithms for segmenting out individual characters in a given image of a cropped word.