Daniel Handwriting Recognition
Fine-tuning a handwriting recognition (HTR) model on Daniel's handwriting: data collection plan, ground truth, training
About This Project
This project establishes a systematic pipeline to fine-tune the Microsoft TrOCR model for personalized handwriting recognition. It utilizes a structured data collection protocol where users transcribe provided text onto physical sheets, which are then photographed and processed to create a ground-truth dataset. By leveraging a pre-trained base model, the system requires only a few hundred lines of handwritten input to adapt effectively to a specific writing style.
The project incorporates Typst for generating standardized writing templates and outlines a complete workflow for image preprocessing, line segmentation, and model training. It is designed for individuals seeking to automate the digitization of personal notes or journals, providing a reproducible method to achieve high transcription accuracy with minimal manual labeling.