2023-06

Accurate Handwriting Transcription Using Deep Learning

CRNN models for handwriting text recognition trained on GCP. Achieved 12% word accuracy improvement using word beam search and CTC loss.

PythonTensorFlowGoogle Cloud Platform

Overview

A deep learning system for recognizing handwritten text, built and trained collaboratively as a team of three. Explored CRNN (Convolutional Recurrent Neural Network) architectures with CTC loss for sequence-to-sequence transcription.

Results

  • Improved word accuracy by 12% over baseline
  • Reduced character error rate via word beam search decoding
  • Trained and evaluated on Google Cloud Platform

Highlights

  • Trained, wrote about, and presented the model and results
  • Co-authored a paper documenting the approach and findings