---
license: mit
datasets:
- felipebandeira/driverlicenses2k
language:
- en
metrics:
- accuracy
pipeline_tag: image-to-text
---

This model extracts information from EU driver's licenses and returns it as JSON. For optimal performance, we recommend that input images:
- have a size of 1192x772
- have high resolution and do not contain light reflection effects

Accuracy 
- on validation set: 98%
- on set of real licenses: 63.93%

Article describing model:
https://medium.com/@ofelipebandeira/transformers-vs-ocr-who-can-read-better-192e6b044dd3

Article describing synthetic dataset used in training:
https://python.plainenglish.io/how-to-create-synthetic-datasets-of-document-images-5f140dee5e40