Aurora OCRrora.es

Character recognition for difficult real-world documents

Aurora OCR is a proprietary character-recognition and document-understanding engine for Latin and Cyrillic documents: scans, photos and PDFs. It is the shared core of several document and HR-technology products.

Current model

Architecture
ConvNeXt-Small, ImageNet-pretrained, fine-tuned in PyTorch
Parameters
≈ 49.5 M
Character classes
110 Latin and Cyrillic letters and digits
Training data
≈ 1.1 M glyph samples, each validated by a person
Clean characters
≈ 99.85 %
Scan-like characters
≈ 99.65 %
Test set
characters from 76 real PDF résumés not used in training

How accuracy is counted. Upper- and lower-case letters that share one shape (c/C, o/O, s/S and similar) are scored as correct by the character model. Their case is decided afterwards by a separate rule from the geometry of the text line.

What we are working on

Reaching at least 99.8 % under realistic degradation: displaced characters, background artefacts, blur, low contrast, varying stroke weight and phone photos of paper. The training pipeline uses hard-example mining, targeted retraining and simulated scanner and camera damage.

The next stage is line-level OCR.

Products built on the engine

  • Resume MinerRésumé creation and processing for individuals.
  • Resume-to-XMLStructured résumé data for ATS and HR-technology vendors.
  • Fixic PDFEditing of complex PDF documents.
  • Fixic ScanEditing and reflow of scanned documents.
  • Font recognitionIdentifying the typeface of a document.

All products are in closed development.