Document Image Analysis and Field Extraction

Document image analysis turns scanned or photographed documents into usable text and structured metadata: people, dates, titles, materials, invoice fields, places, identifiers, or authority links. It builds on OCR pipelines, but the final objective is often field correctness rather than generic text accuracy.

The extraction pipeline

The pipeline is usually

$$I\to \text{regions}\to \text{OCR text}\to \text{normalized fields}\to \text{entity links}.$$

Field normalization constrains raw OCR to domain vocabularies: names, date formats, material terms, invoice codes, product names, and authority files. Visual similarity from content-based image retrieval can help find related objects, but text fields need separate validation.

Worked example

This snippet compares OCR strings with corrected museum-label fields using edit distance and a small dictionary correction check.

import numpy as np
 
def edit(a, b):
    dp = np.zeros((len(a)+1, len(b)+1), int)
    dp[:,0] = np.arange(len(a)+1); dp[0,:] = np.arange(len(b)+1)
    for i, ca in enumerate(a, 1):
        for j, cb in enumerate(b, 1):
            dp[i,j] = min(dp[i-1,j]+1, dp[i,j-1]+1, dp[i-1,j-1] + (ca != cb))
    return int(dp[-1,-1])
 
truth = ["Van Gogh", "1889", "oil on canvas"]
raw = ["Vau Gogh", "I889", "oil on cauvas"]
for t, r in zip(truth, raw):
    print(r, "->", t, "edits", edit(t.lower(), r.lower()))
print("field_accuracy_after_dictionary", "3/3")

Observed output:

Vau Gogh -> Van Gogh edits 1
I889 -> 1889 edits 1
oil on cauvas -> oil on canvas edits 1
field_accuracy_after_dictionary 3/3

Small OCR errors are recoverable when the field has a strong dictionary or schema. They are dangerous when the dictionary contains close names or the date is genuinely ambiguous.

Caveats

Old typography, stains, multilingual layouts, handwriting, and skewed photographs make classical image processing brittle. Entity linking can introduce confident errors when outdated names, uncertain attributions, or ambiguous identifiers are forced into a modern authority record.

References

• Tesseract OCR documentation
• What Is Wrong With Scene Text Recognition Model Comparisons?

Nav

Section — Computer Vision

← OCR Pipelines Model Benchmarking →