01 / ACADEMIC
Research · accepted at ICEECIT 2026 · ongoing
FormReader — Extracting Data from Handwritten Production Forms
A hybrid OCR system that turns handwritten Turkish production forms into structured, reviewable data. It is the subject of my M.Sc. research at Sakarya University.
- Scanned formhandwritten · Turkish
- Geometric normalizationtemplate-first
- Cell extractioncanonical cells
- RecognitionSVTR + CTC · VLM for free text
- Quality layerauto-accept · human review
- Structured dataExcel · JSON
PROBLEM
In manufacturing plants, downtime durations, downtime causes and scrap counts are still written onto paper forms by hand and retyped into a spreadsheet later. Because these values feed downtime and OEE calculations, a wrong value written silently to a database costs more than a rejected one.
SOLUTION
I built a hybrid pipeline that routes each field to a recognition path suited to its type and attaches a quality decision to every output. The form is first aligned to its template and fields are read from their own cells; a quality layer then sends each value either to automatic acceptance or to a human review queue. The system does not promise guaranteed accuracy; it measures its uncertainty.
- Python
- PaddleOCR
- SVTR + CTC
- OpenCV
- Vision-Language Models
- QLoRA
- React
- .NET
- PostgreSQL
