Surya is a document OCR toolkit that does:

• OCR in 90+ languages that benchmarks favorably vs cloud services
• Line-level text detection in any language
• Layout analysis (table, image, header, etc detection)
• Reading order detection
• Table recognition (detecting rows/columns)

It works on a range of documents (see usage and benchmarks for more details).

Detection	OCR

Layout	Reading Order

Table Recognition

Surya is named for the Hindu sun god, who has universal vision.

Discord is where we discuss future development.

There is a hosted API for all surya models available here:

• Works with PDF, images, word docs, and powerpoints
• Consistent speed, with no latency spikes
• High reliability and uptime

I want surya to be as widely accessible as possible, while still funding my development/training costs. Research and personal usage is always okay, but there are some restrictions on commercial usage.

The weights for the models 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 AND under $5M in lifetime VC/angel funding raised. You also must not be competitive with the Datalab API. If you want to remove the GPL license requirements (dual-license) and/or use the weights commercially over the revenue limit, check out the options here.

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.

I've included a streamlit app that lets you interactively try Surya on images or PDF files. Run it with:

pip install streamlit
surya_gui

This command will write out a json file with the detected text and bboxes:

surya_ocr DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --langs is an optional (but recommended) argument that specifies the language(s) to use for OCR. You can comma separate multiple languages. Use the language name or two-letter ISO code from here. Surya supports the 90+ languages found in surya/languages.py.
• --lang_file if you want to use a different language for different PDFs/images, you can optionally specify languages in a file. The format is a JSON dict with the keys being filenames and the values as a list, like {"file1.pdf": ["en", "hi"], "file2.pdf": ["en"]}.
• --images will save images of the pages and detected text lines (optional)
• --results_dir specifies the directory to save results to instead of the default
• --max specifies the maximum number of pages to process if you don't want to process everything
• --start_page specifies the page number to start processing from

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• text_lines - the detected text and bounding boxes for each line
  • text - the text in the line
  • confidence - the confidence of the model in the detected text (0-1)
  • polygon - the polygon for the text line in (x1, y1), (x2, y2), (x3, y3), (x4, y4) format. The points are in clockwise order from the top left.
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
• languages - the languages specified for the page
• page - the page number in the file
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Setting the RECOGNITION_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 40MB of VRAM, so very high batch sizes are possible. The default is a batch size 512, which will use about 20GB of VRAM. Depending on your CPU core count, it may help, too - the default CPU batch size is 32.

from PIL import Image
from surya.ocr import run_ocr
from surya.model.detection.model import load_model as load_det_model, load_processor as load_det_processor
from surya.model.recognition.model import load_model as load_rec_model
from surya.model.recognition.processor import load_processor as load_rec_processor

image = Image.open(IMAGE_PATH)
langs = ["en"] # Replace with your languages - optional but recommended
det_processor, det_model = load_det_processor(), load_det_model()
rec_model, rec_processor = load_rec_model(), load_rec_processor()

predictions = run_ocr([image], [langs], det_model, det_processor, rec_model, rec_processor)

The OCR model can be compiled to get an ~15% speedup in total inference time. The first run will be slow while it compiles, though. First set RECOGNITION_STATIC_CACHE=true, then:

import torch

rec_model.decoder.model = torch.compile(rec_model.decoder.model)

This command will write out a json file with the detected bboxes.

surya_detect DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected text lines (optional)
• --max specifies the maximum number of pages to process if you don't want to process everything
• --results_dir specifies the directory to save results to instead of the default

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• bboxes - detected bounding boxes for text
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
  • polygon - the polygon for the text line in (x1, y1), (x2, y2), (x3, y3), (x4, y4) format. The points are in clockwise order from the top left.
  • confidence - the confidence of the model in the detected text (0-1)
• vertical_lines - vertical lines detected in the document
  • bbox - the axis-aligned line coordinates.
• page - the page number in the file
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Setting the DETECTOR_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 440MB of VRAM, so very high batch sizes are possible. The default is a batch size 36, which will use about 16GB of VRAM. Depending on your CPU core count, it might help, too - the default CPU batch size is 6.

from PIL import Image
from surya.detection import batch_text_detection
from surya.model.detection.model 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_text_detection([image], model, processor)

This command will write out a json file with the detected layout.

surya_layout DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected text lines (optional)
• --max specifies the maximum number of pages to process if you don't want to process everything
• --results_dir specifies the directory to save results to instead of the default

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• bboxes - detected bounding boxes for text
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
  • polygon - the polygon for the text line in (x1, y1), (x2, y2), (x3, y3), (x4, y4) format. The points are in clockwise order from the top left.
  • confidence - the confidence of the model in the detected text (0-1). This is currently not very reliable.
  • label - the label for the bbox. One of Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Figure, Section-header, Table, Text, Title.
• page - the page number in the file
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Setting the DETECTOR_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 400MB of VRAM, so very high batch sizes are possible. The default is a batch size 36, which will use about 16GB of VRAM. Depending on your CPU core count, it might help, too - the default CPU batch size is 6.

from PIL import Image
from surya.detection import batch_text_detection
from surya.layout import batch_layout_detection
from surya.model.detection.model import load_model, load_processor
from surya.settings import settings

image = Image.open(IMAGE_PATH)
model = load_model(checkpoint=settings.LAYOUT_MODEL_CHECKPOINT)
processor = load_processor(checkpoint=settings.LAYOUT_MODEL_CHECKPOINT)
det_model = load_model()
det_processor = load_processor()

# layout_predictions is a list of dicts, one per image
line_predictions = batch_text_detection([image], det_model, det_processor)
layout_predictions = batch_layout_detection([image], model, processor, line_predictions)

This command will write out a json file with the detected reading order and layout.

surya_order DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected text lines (optional)
• --max specifies the maximum number of pages to process if you don't want to process everything
• --results_dir specifies the directory to save results to instead of the default

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• bboxes - detected bounding boxes for text
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
  • position - the position in the reading order of the bbox, starting from 0.
  • label - the label for the bbox. See the layout section of the documentation for a list of potential labels.
• page - the page number in the file
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Setting the ORDER_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 360MB of VRAM, so very high batch sizes are possible. The default is a batch size 32, which will use about 11GB of VRAM. Depending on your CPU core count, it might help, too - the default CPU batch size is 4.

from PIL import Image
from surya.ordering import batch_ordering
from surya.model.ordering.processor import load_processor
from surya.model.ordering.model import load_model

image = Image.open(IMAGE_PATH)
# bboxes should be a list of lists with layout bboxes for the image in [x1,y1,x2,y2] format
# You can get this from the layout model, see above for usage
bboxes = [bbox1, bbox2, ...]

model = load_model()
processor = load_processor()

# order_predictions will be a list of dicts, one per image
order_predictions = batch_ordering([image], [bboxes], model, processor)

This command will write out a json file with the detected table cells and row/column ids, along with row/column bounding boxes.

surya_table DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected table cells + rows and columns (optional)
• --max specifies the maximum number of pages to process if you don't want to process everything
• --results_dir specifies the directory to save results to instead of the default
• --detect_boxes specifies if cells should be detected. By default, they're pulled out of the PDF, but this is not always possible.
• --skip_table_detection tells table recognition not to detect tables first. Use this if your image is already cropped to a table.

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• cells - detected table cells
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
  • row_id - the id of the row this cell belongs to.
  • col_id - the id of the column this cell belongs to.
  • text - if text could be pulled out of the pdf, the text of this cell.
• rows - detected table rows
  • bbox - the bounding box of the table row
  • row_id - the id of the row
• cols - detected table columns
  • bbox - the bounding box of the table column
  • col_id- the id of the column
• page - the page number in the file
• table_idx - the index of the table on the page (sorted in vertical order)
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Setting the TABLE_REC_BATCH_SIZE env var properly will make a big difference when using a GPU. Each batch item will use 150MB of VRAM, so very high batch sizes are possible. The default is a batch size 64, which will use about 10GB of VRAM. Depending on your CPU core count, it might help, too - the default CPU batch size is 8.

• This is specialized for document OCR. It will likely not work on photos or other images.
• It is for printed text, not handwriting (though it may work on some handwriting).
• The text detection model has trained itself to ignore advertisements.
• You can find language support for OCR in surya/languages.py. Text detection, layout analysis, and reading order will work with any language.

If OCR isn't working properly:

• Try increasing resolution of the image so the text is bigger. If the resolution is already very high, try decreasing it to no more than a 2048px width.
• Preprocessing the image (binarizing, deskewing, etc) can help with very old/blurry images.
• You can adjust DETECTOR_BLANK_THRESHOLD and DETECTOR_TEXT_THRESHOLD if you don't get good results. DETECTOR_BLANK_THRESHOLD controls the space between lines - any prediction below this number will be considered blank space. DETECTOR_TEXT_THRESHOLD controls how text is joined - any number above this is considered text. DETECTOR_TEXT_THRESHOLD should always be higher than DETECTOR_BLANK_THRESHOLD, and both should be in the 0-1 range. Looking at the heatmap from the debug output of the detector can tell you how to adjust these (if you see faint things that look like boxes, lower the thresholds, and if you see bboxes being joined together, raise the thresholds).

If you want to develop surya, you can install it manually:

• git clone https://github.com/VikParuchuri/surya.git
• cd surya
• poetry install - installs main and dev dependencies
• poetry shell - activates the virtual environment

Full language results

Tesseract is CPU-based, and surya is CPU or GPU. I tried to cost-match the resources used, so I used a 1xA6000 (48GB VRAM) for surya, and 28 CPU cores for Tesseract (same price on Lambda Labs/DigitalOcean).

I benchmarked OCR against Google Cloud vision since it has similar language coverage to Surya.

Full language results

Methodology

I measured normalized sentence similarity (0-1, higher is better) based on a set of real-world and synthetic pdfs. I sampled PDFs from common crawl, then filtered out the ones with bad OCR. I couldn't find PDFs for some languages, so I also generated simple synthetic PDFs for those.

I used the reference line bboxes from the PDFs with both tesseract and surya, to just evaluate the OCR quality.

For Google Cloud, I aligned the output from Google Cloud with the ground truth. I had to skip RTL languages since they didn't align well.

Tesseract is CPU-based, and surya is CPU or GPU. I ran the benchmarks on a system with an A10 GPU, and a 32 core CPU. This was the resource usage:

• tesseract - 32 CPU cores, or 8 workers using 4 cores each
• surya - 36 batch size, for 16GB VRAM usage

Methodology

Surya predicts line-level bboxes, while tesseract and others predict word-level or character-level. It's 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 the 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.

Time per image - .4 seconds on GPU (A10).

Methodology

I benchmarked the layout analysis on Publaynet, which was not in the training data. I had to align publaynet labels with the surya layout labels. I was then able to find coverage for each layout type:

• Precision - how well the predicted bboxes cover ground truth bboxes
• Recall - how well ground truth bboxes cover predicted bboxes

75% mean accuracy, and .14 seconds per image on an A6000 GPU. See methodology for notes - this benchmark is not perfect measure of accuracy, and is more useful as a sanity check.

Methodology

I benchmarked the reading order on the layout dataset from here, which was not in the training data. Unfortunately, this dataset is fairly noisy, and not all the labels are correct. It was very hard to find a dataset annotated with reading order and also layout information. I wanted to avoid using a cloud service for the ground truth.

The accuracy is computed by finding if each pair of layout boxes is in the correct order, then taking the % that are correct.

Higher is better for intersection, which the percentage of the actual row/column overlapped by the predictions.

Methodology

The benchmark uses a subset of Fintabnet from IBM. It has labeled rows and columns. After table recognition is run, the predicted rows and columns are compared to the ground truth. There is an additional penalty for predicting too many or too few rows/columns.

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

• --max controls how many images to process for the benchmark
• --debug will render images and detected bboxes
• --pdf_path will let you specify a pdf to benchmark instead of the default data
• --results_dir will let you specify a directory to save results to instead of the default one

Text recognition

This will evaluate surya and optionally tesseract on multilingual pdfs from common crawl (with synthetic data for missing languages).

python benchmark/recognition.py --tesseract

• --max controls how many images to process for the benchmark

• --debug 2 will render images with detected text

• --results_dir will let you specify a directory to save results to instead of the default one

• --tesseract will run the benchmark with tesseract. You have to run sudo apt-get install tesseract-ocr-all to install all tesseract data, and set TESSDATA_PREFIX to the path to the tesseract data folder.

• Set RECOGNITION_BATCH_SIZE=864 to use the same batch size as the benchmark.

• Set RECOGNITION_BENCH_DATASET_NAME=vikp/rec_bench_hist to use the historical document data for benchmarking. This data comes from the tapuscorpus.

Layout analysis

This will evaluate surya on the publaynet dataset.

python benchmark/layout.py

• --max controls how many images to process for the benchmark
• --debug will render images with detected text
• --results_dir will let you specify a directory to save results to instead of the default one

Reading Order

python benchmark/ordering.py

• --max controls how many images to process for the benchmark
• --debug will render images with detected text
• --results_dir will let you specify a directory to save results to instead of the default one

Table Recognition

python benchmark/table_recognition.py --max 1024 --tatr

• --max controls how many images to process for the benchmark
• --debug will render images with detected text
• --results_dir will let you specify a directory to save results to instead of the default one
• --tatr specifies whether to also run table transformer

Text detection was trained on 4x A6000s for 3 days. It used a diverse set of images as training data. It was trained from scratch using a modified efficientvit architecture for semantic segmentation.

Text recognition was trained on 4x A6000s for 2 weeks. It was trained using a modified donut model (GQA, MoE layer, UTF-16 decoding, layer config changes).

This work would not have been possible without amazing open source AI work:

• Segformer from NVIDIA
• EfficientViT from MIT
• timm from Ross Wightman
• Donut from Naver
• transformers from huggingface
• CRAFT, a great scene text detection model

Thank you to everyone who makes open source AI possible.