Surya is a multilingual document OCR toolkit. It can do:
- Accurate line-level text detection
- Text recognition (coming soon)
- Table and chart detection (coming soon)
It works on a range of documents and languages (see usage and benchmarks for more details).
Surya is named after the Hindu sun god, who has universal vision.
Discord is where we discuss future development.
| Name | Text Detection |
|---|---|
| New York Times | Image |
| Japanese | Image |
| Chinese | Image |
| Hindi | Image |
| Presentation | Image |
| Scientific Paper | Image |
| Scanned Document | Image |
You'll need python 3.9+ and PyTorch. You may need to install the CPU version of torch first if you're not using a Mac or a GPU machine. See here for more details.
Install with:
pip install surya-ocr
Model weights will automatically download the first time you run surya.
- Inspect the settings in
surya/settings.py. You can override any settings with environment variables. - Your torch device will be automatically detected, but you can override this. For example,
TORCH_DEVICE=cuda. Note that thempsdevice has a bug (on the Apple side) that may prevent it from working properly.
You can detect text lines in an image, pdf, or folder of images/pdfs with the following command. This will write out a json file with the detected bboxes, and optionally save images of the pages with the bboxes.
surya_detect DATA_PATH --images
DATA_PATHcan be an image, pdf, or folder of images/pdfs--imageswill save images of the pages and detected text lines (optional)--maxspecifies the maximum number of pages to process if you don't want to process everything--results_dirspecifies the directory to save results to instead of the default
This has worked with every language I've tried. It will work best with documents, and may not work well with photos or other images. It will also not work well with handwriting.
Performance tips
Setting the DETECTOR_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 280MB of VRAM, so very high batch sizes are possible. The default is a batch size 32, which will use about 9GB of VRAM.
Depending on your CPU core count, DETECTOR_BATCH_SIZE might make a difference there too - the default CPU batch size is 2.
You can adjust DETECTOR_NMS_THRESHOLD and DETECTOR_TEXT_THRESHOLD if you don't get good results. Try lowering them to detect more text, and vice versa.
You can also do text detection from code with:
from PIL import Image
from surya.detection import batch_inference
from surya.model.segformer import load_model, load_processor
image = Image.open(IMAGE_PATH)
model, processor = load_model(), load_processor()
# predictions is a list of dicts, one per image
predictions = batch_inference([image], model, processor)
Coming soon.
Coming soon.
If you want to develop surya, you can install it manually:
git clone https://github.com/VikParuchuri/surya.gitcd suryapoetry install# Installs main and dev dependencies
- This is specialized for document OCR. It will likely not work on photos or other images. It will also not work on handwritten text.
- Does not work well with images that look like ads or other parts of documents that are usually ignored.
| Model | Time (s) | Time per page (s) | precision | recall |
|---|---|---|---|---|
| surya | 52.6892 | 0.205817 | 0.844426 | 0.937818 |
| tesseract | 74.4546 | 0.290838 | 0.631498 | 0.997694 |
Tesseract is CPU-based, and surya is CPU or GPU. I ran the benchmarks on a system with an A6000 GPU, and a 32 core CPU. This was the resource usage:
- tesseract - 32 CPU cores, or 8 workers using 4 cores each
- surya - 32 batch size, for 9GB VRAM usage
Methodology
Surya predicts line-level bboxes, while tesseract and others predict word-level or character-level. It's also hard to find 100% correct datasets with line-level annotations. Merging bboxes can be noisy, so I chose not to use IoU as the metric for evaluation.
I instead used coverage, which calculates:
- Precision - how well predicted bboxes cover ground truth bboxes
- Recall - how well ground truth bboxes cover predicted bboxes
First calculate coverage for each bbox, then add a small penalty for double coverage, since we want the detection to have non-overlapping bboxes. Anything with a coverage of 0.5 or higher is considered a match.
Then we calculate precision and recall for the whole dataset.
You can benchmark the performance of surya on your machine.
- Follow the manual install instructions above.
poetry install --group dev# Installs dev dependencies
Text line detection
This will evaluate tesseract and surya for text line detection across a randomly sampled set of images from doclaynet.
python benchmark/detection.py --max 256
--maxcontrols how many images to process for the benchmark--debugwill render images and detected bboxes--pdf_pathwill let you specify a pdf to benchmark instead of the default data--results_dirwill let you specify a directory to save results to instead of the default one
This was trained on 4x A6000s for about 3 days. It used a diverse set of images as training data. It was trained from scratch using a modified segformer architecture that reduces inference RAM requirements.
Text detection
The text detection model was trained from scratch, so it's okay for commercial usage. The weights are licensed cc-by-nc-sa-4.0, but I will waive that for any organization under $5M USD in gross revenue in the most recent 12-month period.
If you want to remove the GPL license requirements for inference or use the weights commercially over the revenue limit, please contact me at surya@vikas.sh for dual licensing.
This work would not have been possible without amazing open source AI work:
- Segformer from NVIDIA
- transformers from huggingface
- CRAFT, a great scene text detection model
Thank you to everyone who makes open source AI possible.