surya-ocr

• Move cell assignment logic into a separate library I'm creating, tabled
• Improve cell extraction from PDFs

• Small bugfix after the table recognition release

• Add a new table recognition model that detects rows/columns and cells
• Add benchmarks for accuracy and speed (seems to be very accurate wrt to current state of the art open model)
• Improve memory efficiency of layout and text detection (hopefully no more memory leaks)
• Improve resolution handling for layout/text detection/ocr, which should improve accuracy quite a bit

A new version of the OCR model with a custom architecture.

• 20% faster
• Automatic language detection, with support for optional language hints
• Better accuracy on old/noisy documents
• Basic english handwriting support (to be improved soon)

Switched model architecture for the text detection and layout models:

• 30% faster on GPU
• 4x faster on CPU
• 12x faster on MPS (M series macs)

Accuracy should be about the same, or slightly better, from my benchmarks.

New transformers version added a new kwarg to donut embeddings. This now handles and ignores that kwarg, and also slightly future-proofs in case this happens again.

• Fix bugs with RGBA images
• Fix assert bug
• Add back in thumbnail method for resizing
• Slightly optimize segformer code

• Image resize from cv2 to PIL - cv2 caused benchmark regressions

• Speed up base OCR model ~15-20%, and reduce memory usage by ~25% (can do higher batch sizes)
• Add static cache for compilation - torch.compile will result in another 15% speedup
• Other optimizations, like faster image resizing
• Bugfixes, like enabling different length language inputs for OCR (batching different docs with different languages together)

• Remove unneeded format conversions
• Fix bug in OCR, where only one color channel was used for OCR - results should be better now
• Speed up layout/text detection a bit

Cut OCR time in half. Combined with the previous release, OCR should now take about 40% as much time as it did before.

• Improve CPU postprocessing for line detection and layout - cut postprocessing time to 1/3 of original
• Unpin transformers version after investigating model performance

This should result in an ~2x speedup for layout and text detection. The effect will be most noticeable on GPU. I haven't fully benchmarked, though.

• Fix memory leak with layout and text detection models and large batch sizes
• Improve ordering model generation slightly

• Prune MoE experts before loading model
• Unpin torch version from 2.2.2

• Programmatic batch sizes for all models
• Update package versions
• Bugfixes

• Release reading order model
• Add Google Cloud OCR benchmarks to README

• Add layout model
• Multiprocessing for detection postprocessing (can be much faster)
• Fix minor bugs

• Improved line detector with better recall and higher resolution
• Confidence scores for detection and recognition
• Beta layout detection model

• Output JSON as UTF-8 so text can be scanned easily
• Improve output schema to include more information and be easier to parse
• Sort text lines in basic order, which will work in many cases
• Detect repetitions and truncate

• Fix bug with downloading render fonts for OCR

• Add in text recognition
• Add simple streamlit demo app

• Much more performant on scanned/old docs
• Enable non-axis-aligned bboxes

Initial surya text detection release.