The goal was to create an OCR module that adapted to the material it was given. I built a grid search that tunes the module to each document set, trying every parameter combination across repeated trials until the best set emerges. Most OCR pipelines fix one configuration and accept the accuracy it gives. This one fits itself to the source, which matters when the text feeds a model downstream and errors propagate. Grid search is brute force, but it worked.
Built with
- Machine Learning
- Optical Character Recognition
- Python