Inspiration
At MRG Laboratories, researchers manually transcribe handwritten information from cylindrical grease sample tubes into spreadsheets. We wanted to eliminate that repetitive work so researchers could focus on actual analysis.
Technology
GreaseFreeData uses a Python-based computer vision pipeline to extract handwritten data from rotating sample tubes. Using OpenCV and Laplacian variance for sharpness detection, the system selects eight high-quality frames from each video. These frames are then processed by an open-source OCR model to recognize and reconstruct information across curved labels.
Our updated pipeline targets approximately 6 seconds per sample with over 95% recognition accuracy, based on preliminary testing, while eliminating dependence on paid cloud-based AI APIs.
The platform combines React, TypeScript, FastAPI, PostgreSQL, Redis, and Docker for scalable batch processing, persistent storage, and human verification.
Design
We designed an intuitive workflow: Scan → Review → Verify → Export. Researchers can batch-upload videos, inspect extracted data alongside original footage, correct suspicious fields highlighted in red, and selectively export records to CSV.
Completion
We built a functional end-to-end platform featuring batch uploads, automated extraction, editable records, date anomaly detection, human verification, and spreadsheet exports. Our latest improvement replaces cloud-based inference with local, open-source OCR, which is currently being integrated.
Learning
Our biggest challenge was reconstructing handwritten information spread across multiple frames of a rotating cylindrical label. We explored computer vision, image sharpness optimization, OCR, asynchronous processing, and database engineering while learning to balance speed, accuracy, reliability, and usability.
Built With
- docker
- fastapi
- javascript
- laplaciansharpening
- ocr
- opencv
- postgresql
- python
- react
- redis
- typescript
- vite
