Arsalan Younus.
Back to Projects

OCR Text Recognition

Domain-tuned text recognition for both printed and handwritten content, removing recognition accuracy as a bottleneck for the end-to-end pipeline.

The Business Problem

After localizing word regions, off-the-shelf OCR performed poorly on handwriting and mixed print-and-handwriting documents. Low recognition accuracy became a bottleneck for the entire extraction pipeline.

The client needed domain-tuned recognition that handles both printed and handwritten text from cropped regions reliably.

The Technical Solution

I built an end-to-end text recognition system using CRNN and Transformer-based architectures (trOCR), trained on domain data for both printed and handwritten text.

The system takes cropped word or line images from the localization step and outputs recognized text as a reliable plug-in for the OCR stack.

The Scalability Factor

Deployed on AWS with Docker alongside localization and post-processing stages. Containerized inference supports batch processing at production document volumes.

Business Impact

Recognition accuracy reached 94% on mixed printed and handwritten content.

The recognition stage is a reliable plug-in deployed alongside localization and post-processing.

Built with

CRNN
trOCR
Transformers
PyTorch
OpenCV
AWS
Docker
OCR Text Recognition screenshot 1
View