OCR data extraction: how it works, methods and accuracy checks

OCR data extraction reads the text on a scan, photo or PDF and turns it into named fields and tables. The first step is OCR, short for optical character recognition, which reads the characters on the page. Extraction then pulls out values such as an invoice number, a closing balance or a policy date. Extraction also identifies what each value is, checks the value and sends it to the system that uses it.

What is OCR data extraction?#

OCR data extraction, also called OCR extraction or OCR data capture, has two jobs. OCR turns an image of a page into text, but the text has no labels. OCR doesn't know that "4,120.50" is the invoice total, or that a column of numbers belongs to a table. Extraction adds that knowledge. Document layout analysis finds the parts of the page, such as the header, the table and the totals. It puts them in the order a person would read them. Field extraction then names each value and checks it.

The output is structured data, such as JSON, a spreadsheet row or a record in your system, not a searchable file. A digital PDF already carries a text layer, so it can skip OCR. Scans, faxes and phone photos can't.

How OCR data extraction works#

An OCR document processing workflow takes a page image through these stages until it becomes a record in your system.

Poor scans are straightened and cleaned up first, because OCR errors pass on to every later step. A misread digit becomes a wrong amount in the extracted data. Extraction returns each field and table row, usually with a confidence score, which shows how sure the software is about that value. Validation checks the values against rules. Only values that fail a check, or that score below the confidence threshold you set, go to a person. Here is the extraction stage on a business bank statement.

OCR data extraction methods#

OCR-based extraction tools use four methods, often in combination. What separates them is how each copes with a layout it hasn't seen before.

• Zonal or template OCR

  Reads fixed boxes on a known layout. Fast and cheap for one form version; breaks when a vendor moves a field or a scan shifts. See zonal OCR.

• Keywords and regular expressions

  Finds a label such as "Invoice No." and the value after it, using a fixed text pattern called a regular expression. Works across layouts, but fails on unusual labels, values that wrap onto a second line and other numbers that fit the same pattern.

• Table extraction

  Rebuilds rows and columns from the position of each word. Needed for line items and transactions; breaks on merged cells, missing ruling lines and tables that run across pages.

• AI field extraction

  A model trained on the document type finds each field from the text around it and its position on the page, with a confidence score. Handles new layouts, but needs a review step for low-confidence values.

Handwriting and checkboxes need their own readers. Intelligent character recognition (ICR) reads hand-printed text, and optical mark recognition (OMR) reads marks. Free, open-source tools such as Tesseract return the text and where each word sits on the page. The output comes in formats such as plain text, hOCR or TSV. If you build your own system on one of these tools, it must still extract fields, check values and send doubtful ones to a reviewer.

What documents can OCR extract data from?#

Any document with text on it. The payoff is biggest on documents that arrive in volume and in many layouts.

• Invoices and receipts

  Vendor, invoice number, dates, line items, tax and total. See invoice extraction.

• Bank statements

  Account details, balances and every transaction, across banks and formats. See bank statement extraction.

• Pay stubs

  Employer, pay period, gross and net pay, for the period and year to date. See pay stub extraction.

• Tax forms

  W-2s, 1099s and returns, box by box, with mixed packets sorted by form. See IRS tax form extraction.

• ACORD forms

  Insurers, policy numbers, limits and dates from certificates and applications. See ACORD form processing.

• IDs

  Name, date of birth, document number and expiry from passports and driver's licenses. See passport OCR and driver's license OCR.

How to extract data from scanned documents and images#

For a few documents a month, copying values by hand is fine. At hundreds or thousands, set up a pipeline. Take one paper invoice that someone photographed and sent by email.

The attachment goes into the same queue as uploads and scanner output, and the original image stays with the record. The software finds no text layer, so the photo goes to OCR. A digital PDF would skip that step. If a page were cut off or unreadable, someone would re-request it. Classification then labels the file as an invoice, so it reaches the invoice model. A combined PDF, such as a loan packet, would be split first so that each page reaches the right model.

OCR returns the text with no labels, and layout analysis finds the header, the line items and the totals. Extraction names the vendor, the invoice number and the dates, and it rebuilds the table row by row. If the table continues across pages, extraction joins the pages into one table. Validation then adds the line items and the tax and compares the sum with the 4,120.50 printed as the total. It also checks dates, formats and required fields, and where the file holds other documents, it compares values across them.

If everything agrees, the data goes to the accounting system through an API. If a blurred digit makes the sum and the total disagree, the field the software was unsure of goes to a person. That person checks it against the original image, and the other fields need no review.

Play video: A 5-minute guide to OCR software

Manual data entry vs OCR data extraction#

Accounts payable shows the difference clearly, because invoices arrive in every format.

Docsumo is an intelligent document processing (IDP) platform. It reads printed and handwritten text in scans, PDFs and phone photos. It extracts named fields and multi-page tables, checks them and sends only the fields it isn't sure about to a person. Docsumo reports these results.

• 99%99%field-level accuracy across 250+ document types
• 95%+95%+of documents processed straight through, without manual review
• <5 min<5 minper document, down from 2+ hours

How accurate is OCR data extraction?#

Accuracy is measured at two levels. Character accuracy counts the letters and digits OCR reads correctly. Field-level accuracy counts the extracted values that are right. It is the one that matters to your systems, because one wrong digit spoils a whole amount. Character accuracy drops with poor scans (low resolution, skew, glare) and with handwriting. Field-level accuracy also drops with layouts the tool hasn't seen. OCR accuracy covers how to measure both.

We'd rather count corrections than quote a percentage. Put a batch of your own documents through the tool, poor scans included, and count the fields a reviewer had to fix. That count reflects your layouts and your worst files, so we'd trust it over any headline accuracy figure, ours included.

Whatever tool you use, run these checks before extracted data posts.

• Totals reconcileLine items plus tax equal the total; opening balance plus deposits minus withdrawals equals the closing balance.
• Required fields are presentEvery document has the fields your process needs, such as an invoice number and a due date.
• Formats are validDates are real dates, amounts are numbers and IDs match their expected pattern.
• Values match your recordsVendor names, account numbers and PO numbers exist in your master data.
• Documents agreeThe same name, amount or date matches across the documents in one file.
• Uncertain fields get a personSet a confidence threshold per field and route anything below it to review.

Which extracted values can skip review?#

The validation step decides. A value that passes your checks and scores above the confidence threshold can go into your system without a person reviewing it. Everything else goes to a reviewer. OCR turns a page into text, and extraction turns that text into fields and tables you can use. So start with the documents that arrive in volume, measure accuracy at the field level, and send only the exceptions to review.

Book a demo with a few of your own documents, or start a free trial.

Frequently asked questions#