| license: apache-2.0 | |
| library_name: onnx | |
| tags: | |
| - ocr | |
| - doctr | |
| - text-detection | |
| - text-recognition | |
| # docTR ONNX models for OpenMasq | |
| ONNX exports of two pretrained [docTR](https://github.com/mindee/doctr) models by Mindee, used by the OpenMasq desktop app for on-device OCR of Latin-script documents: | |
| | file | docTR architecture | sha256 | | |
| |---|---|---| | |
| | `db_mobilenet_v3_large.onnx` | text detection, `db_mobilenet_v3_large` | `5a82788a1907dccec9978c756f56d386fd2242597ba2630322da063af44cf4d3` | | |
| | `crnn_mobilenet_v3_small.onnx` | text recognition, `crnn_mobilenet_v3_small` | `d89bbd3e732261c341c4cd50e3ad879233c852062866fd55b32c7d431dce301d` | | |
| **Provenance.** Exported from Mindee's official pretrained weights with docTR's own `export_model_to_onnx`; no retraining, no third-party re-upload. | |
| **Integrity.** The app pins these sha256 values: the build verifies each file before bundling it, and the app verifies it again before onnxruntime loads it. A different byte content is refused. | |
| **License.** Apache-2.0, as docTR and its pretrained weights. | |