Refactor surya; new table recognition model

Refactor

This is a complete refactor of surya - the code is now cleaner and better organized. Models are now imported and used differently, here is an example for OCR:

from PIL import Image
from surya.recognition import RecognitionPredictor
from surya.detection import DetectionPredictor

image = Image.open(IMAGE_PATH)
langs = ["en"] # Replace with your languages or pass None (recommended to use None)
recognition_predictor = RecognitionPredictor()
detection_predictor = DetectionPredictor()

predictions = recognition_predictor([image], [langs], detection_predictor)

See the README for how to use other models.

Table recognition

There is a new table recognition model which detects colspans/rowspans better, along with header cells. It also isn't as complex to use, since it operates on just the images versus the images and bboxes.

What's Changed

• Layout improvements by @VikParuchuri in #267
• New table model; total refactor by @VikParuchuri in #279
• Add ci workflow by @VikParuchuri in #284

Full Changelog: v0.8.3...v0.9.0

Pin pypdfium2

Pin pypdfium2 version - newest version can cause issues.

New layout model

Layout model is twice as fast and more accurate.

What's Changed

• Update layout model by @VikParuchuri in #270

Full Changelog: v0.8.1...v0.8.2

Add bad OCR detection model

• Add a model to detect bad OCR text
• Add top_k predictions to layout
• Add in test suite

What's Changed

• Add OCR Error Detection Model by @tarun-menta in #261
• Add top_k to Surya Layout and Fix Confidence Value Issue by @iammosespaulr in #263
• Bad OCR detection model by @VikParuchuri in #268

New Contributors

• @tarun-menta made their first contribution in #261

Full Changelog: v0.8.0...v0.8.1

Surya Bugfixes and Improvements to `pdftext`

Update to the latest pdftext release, incorporating heuristic-based segmentation for enhanced performance and accuracy.

Full Changelog: v0.7.0...v0.8.0

New layout model

• New layout model that detects more block types, includes ordering, and performs better
• Remove ordering model
• Refactor internals to use common model definitions and processors
• Add flags for easier compilation

What's Changed

• Fix detection results being mutated in place by @iammosespaulr in #236
• Add support for compiled inference by @iammosespaulr in #237
• Dev by @VikParuchuri in #243
• Layout2 by @VikParuchuri in #249
• New layout model by @VikParuchuri in #252

Full Changelog: v0.6.13...v0.7.0

Performance Improvements

• Don't pass around heatmaps by default
• Use threadpool vs processpool
• Optimize fn for postprocessing

Flatten table form fields, overlap postprocessing

• Overlap postprocessing if batch size >1
• Flatten in form fields with tables

Transformers 4.46, torch 2.5 fixes

• Fixes cudnn backend issue with torch 2.5
• Fixes issue with sdpa and transformers 4.46
• Flatten form fields into pdf by default for table rec

Threads

Threads cause issues on a small % of devices. Although they do give good speedups on most, supporting them seems like a bad idea.