VietOCR
What VietOCR is
VietOCR is a free Python library for recognizing printed and handwritten Vietnamese text in images. The project provides Attention Seq2Seq and Transformer OCR architectures, along with pretrained VGG Seq2Seq and VGG Transformer models. Its Python Predictor can run inference using pretrained models or user-trained weights, with CUDA or CPU device settings shown. Users can train models on their own datasets and define custom image augmentation. The project says its models were trained on a 10-million-image dataset containing generated images, handwriting and scanned documents. Training and test annotation files pair an image filename with its label, separated by a tab. For its 10-million-image experiment, the README reports full-sequence precision of 0.8800 for VGG Transformer and 0.8701 for VGG Seq2Seq. It reports prediction times of 86 ms for VGG Transformer and 12 ms for VGG Seq2Seq on a 1080 Ti. The author cautions that pretrained models may be sensitive to small input-image changes on new datasets and recommends retraining for practical use. Installation instructions use `pip install vietocr`; the library is released under the Apache 2.0 license.
Who it suits
VietOCR may suit developers working with Vietnamese text in images who want a Python library and the option to train models on their own datasets. Its pretrained models and CPU or CUDA inference settings offer alternatives for running prediction.
What it does well
- Recognizes printed and handwritten Vietnamese text.
- Includes Transformer and Attention Seq2Seq model types.
- Users can train models on their own datasets.
- Apache 2.0 license.
What to know first
- Pretrained models may react to small image changes.
- Author recommends retraining for practical use on new datasets.
- Recognition is for Vietnamese text in images.
Plans
| Plan | Billed and limits | Price |
|---|---|---|
| VietOCR | Open-source Python library · Apache 2.0 license | Free |
As github.com lists them, read 11 Oct 2026. The maker's own page has the final word.
Verdict
VietOCR provides pretrained options and training tools for Vietnamese image OCR, with reported speed and precision figures for its two model choices. Its stated sensitivity to input changes on new datasets makes retraining worth considering for practical use.
Facts from VietOCR's own pages
- Dataset format
- Training and test annotation files contain an image filename and label separated by a tab. · github.com, 11 Oct 2026
- Free plan
- Yes · github.com, 23 Sept 2026
- Handwriting OCR
- Yes · github.com, 23 Sept 2026
- Inference
- The Python Predictor can run inference using a pretrained model or user-trained weights, with CUDA or CPU device settings shown. · pbcquoc.github.io, 11 Oct 2026
- Supported inputs
- image · github.com, 23 Sept 2026
- Installation
- The project README gives `pip install vietocr` as the installation command. · github.com, 11 Oct 2026
- Known limitation
- The author says pretrained models can be sensitive to small input-image changes on new datasets and recommends retraining for practical use. · pbcquoc.github.io, 11 Oct 2026
- License
- The GitHub repository states that the library is released under the Apache 2.0 license. · github.com, 11 Oct 2026
- Model architectures
- It provides Attention Seq2Seq and Transformer OCR model types. · github.com, 11 Oct 2026
- Model data
- The project says its models were trained on a 10-million-image dataset that includes generated images, handwriting, and scanned documents. · github.com, 11 Oct 2026
- Model options
- The project offers VGG Transformer and VGG Seq2Seq pretrained models. · github.com, 11 Oct 2026
- Performance
- The README reports full-sequence precision of 0.8800 for VGG Transformer and 0.8701 for VGG Seq2Seq on its 10-million-image experiment. · github.com, 11 Oct 2026
- Purpose
- VietOCR is a library for OCR, with models for recognizing handwritten and printed Vietnamese text. · github.com, 11 Oct 2026
- Speed tradeoff
- The README reports prediction times of 86 ms for VGG Transformer and 12 ms for VGG Seq2Seq on a 1080 Ti. · github.com, 11 Oct 2026
- Support
- The README directs users with problems to create an issue or contact the author by email. · github.com, 11 Oct 2026
- Training
- Users can train models on their own datasets and define custom image augmentation. · pbcquoc.github.io, 11 Oct 2026
Where VietOCR stands on Freedom251
- OCR software#29 of 40