Handwriting & form understanding accuracy evaluation.

Both evaluations rely on the Gestalt Pattern Matching algorithm (Python's difflib.SequenceMatcher) to score similarity between extracted text and human-annotated ground truth, from 0.0 (completely different) to 1.0 (identical). All comparisons run on case-normalized strings, so partial matches and common OCR errors are credited proportionally rather than all-or-nothing.

Field-specific evaluators handle each data type on its own terms: dates are normalized to ISO format with dual MM/DD vs DD/MM validation to absorb annotation inconsistencies in the dataset, and totals are converted to floats and compared exactly. Blurry and low-quality images were deliberately kept in the sample to reflect real-world conditions, and empty documents were tested for correct null detection.

500 samples of handwritten personal names — script, print, and hybrid styles, including deliberately blurry and low-quality images — were processed through a single-field extraction schema. Axia averaged a 93.2% string-similarity score against human annotations, at 2.0 seconds per document, and correctly returned null for images containing no text.

347 real Malaysian business receipts from SROIE2019 — mixed Malay and English, varied layouts and print quality — were processed with a four-field schema extracting company name, date, address, and total amount, at 4.2 seconds per receipt.