What we measure
A receipt list, paper ledger, or handwritten invoice grid can be readable to a person and still be hard to restore into rows and columns. The benchmark therefore measures recognition, structure recovery, and tolerance for imperfect captures.
Handwriting accuracy
96.8%
on handwritten table samples
Table recovery
99.1%
rows and cell relationships
Noise tolerance
94.7%
blurred, skewed, and low-light pages
Accuracy view
The chart keeps the page readable while still showing the extraction story. Text accuracy matters first, but the benchmark is read next to table fidelity because spreadsheet teams pay for the reconstruction time after OCR, not only for transcribed words.
Based on 10,000+ real-world samples
Tested on IAM Handwriting Database v3.0
Workbook metrics
The comparison table records the signals users feel during review. Lower character error rates reduce spot fixes. Stronger structure recovery reduces reformatting. Noise tolerance matters when the original document came from a desk photo instead of a clean scanner.
| Metric | AxLiner | Industry Avg |
|---|---|---|
Character Error Rate Lower is better | 3.2% | 5.8% |
Word Recognition Handwriting and mixed print | 99.5% | 95.1% |
Table Structure Rows, headers, and cell boundaries | 99.1% | 92.3% |
Noisy Image Handling Phone photos and low contrast scans | 94.7% | 87.2% |
Mixed Font Recognition Printed text plus handwriting | 97.9% | 94.6% |
Processing Speed Median measured page time | 0.8s/page | 2.1s/page |
Review protocol
The final check follows the user flow. The source page is prepared, the model reads the handwriting, the table is reconstructed, and the exported workbook is reviewed where corrections happen.
Normalize
Start from the page actually received: phone capture, scan, screenshot, rotated PDF page, or low-contrast handwritten form.
Read
Check characters and words without forgetting the lines, merged areas, headers, totals, and writing that crosses ruled cells.
Rebuild
Measure whether recovered rows and columns stay useful after export instead of only scoring loose OCR text.
Review
Look at the workbook in the correction flow, because the product promise is an editable spreadsheet batch.