Starstruk Logo MarkStarstruk

v0.11.0

datalab-to/suryav0.11.0Feb 10, 2025by VikParuchuri

AI Summary

Adds inline math detection, benchmarks against Textract/Google Cloud Vision, and introduces XLA support for TPUs.

Key Highlights

  • New inline math detection model.
  • Benchmark shows Surya is faster than Textract for English (0.52s vs 1.44s per page).
  • Adds XLA support for TPUs.
  • Refactored inference for 5-10% speed boost.

New Features

  • Add XLA support
  • Add Inline Math Detection
  • Add Textract OCR Benchmark
  • Update line detection model

Full Release Notes

## Inline math detection

- Add new inline math detection model and benchmark

<img width="464" alt="image" src="https://github.com/user-attachments/assets/3581838f-3bf4-491a-8cd3-b135c0de623a" />

## Textract OCR benchmark

Benchmark surya against textract as well as google cloud vision.  For just english, results look like:

| Model     |   Time per page (s) |   Avg Score |   English |
|----------|---------------------|-------------|-----------|
| surya      |           0.522628  |    0.983298 |  0.983298 |
| textract  |            1.44293 |    0.947458 |  0.947458 |


## XLA support

Add support for TPUs.  Still fairly slow, but lots of optimizations to be made.

## Minor speedups

Refactor inference to get a 5-10% speed boost across all models.

## What's Changed
* Add XLA support by @iammosespaulr in https://github.com/VikParuchuri/surya/pull/298
* Add Inline Math Detection by @tarun-menta in https://github.com/VikParuchuri/surya/pull/297
* Update to new line detection model by @tarun-menta in https://github.com/VikParuchuri/surya/pull/305
* Fix merging of inline boxes by drawing textlines in heatmap by @tarun-menta in https://github.com/VikParuchuri/surya/pull/309
* XLA improvements by @VikParuchuri in https://github.com/VikParuchuri/surya/pull/306
* Update inline math checkpoint by @VikParuchuri in https://github.com/VikParuchuri/surya/pull/310
* Misc Line Detection Fixes by @tarun-menta in https://github.com/VikParuchuri/surya/pull/313
* Add Textract OCR Benchmark by @tarun-menta in https://github.com/VikParuchuri/surya/pull/307
* Inline math model, new text detection model by @VikParuchuri in https://github.com/VikParuchuri/surya/pull/312


**Full Changelog**: https://github.com/VikParuchuri/surya/compare/v0.10.3...v0.11.0