What OCR is: a complete guide to reading real business documents

OCR (Optical Character Recognition) finds the characters a person reads in a document that exists only as an image, such as a scan or a photograph, converts them into text a computer can handle, and returns the position of each character alongside.

Why OCR is still difficult

Clean print is a solved problem.

Accuracy on neatly printed documents reached a usable level a long time ago, which is why OCR is widely assumed to be finished technology. Open an actual intake queue and the picture changes.

Look at one financial institution's received documents. On one, a black band from a page folded during scanning has swallowed half of an authenticity mark. Another has lost margins and part of its body text through repeated photocopying. There are documents photographed on a desk, arriving with creases, shadows, and distortion, and faxes retransmitted four times where four different date formats now sit stacked on a single line. On top of that, stamps intrude into tables, strikethroughs cross characters, and handwriting and checkboxes mix together.

People read documents like these without effort. What is hard in character recognition is not reading print; it is reading documents in that condition.

What OCR actually is: how it differs from text extraction

Text extraction and OCR across three axes

First, the target differs. Text extraction pulls out character data already stored inside a digitally authored document. The characters exist as data in the source file, so it is a matter of reading them. OCR works on an image where that data does not exist. What is on screen is pixels, and the work starts with finding the characters among them.

Second, the objective differs. Text extraction aims to move stored strings across without loss. OCR aims to find every character a person could read in an image and restore it as text. Accuracy is therefore barely a concept in text extraction while it is the central metric in OCR.

Third, the metrics differ. Text extraction is judged on whether anything was dropped. OCR is judged on character-level accuracy together with fuzzy similarity. In real validations, one financial document set confirmed roughly 98 percent on both printed and handwritten fields by fuzzy similarity, with recognition speed measured at 0.5 seconds per page on accelerator hardware. Another handwriting-heavy set confirmed 98.7 percent.

The two are not alternatives. Digital originals and scans coexist within a single workflow, so the configuration has to identify each file's nature and process it accordingly. Converting a document that already has text into an image to recognise it lowers accuracy for no reason.

Six conditions for putting OCR into production

Six conditions that hold in real operations

First, document condition has to be corrected. Skewed scans, rotated pages, and low-resolution faxes need fixing before recognition. Confirm that deskewing, rotation correction, and super-resolution sit in the front of the pipeline.

Second, it has to withstand obstructions. Stamps, watermarks, strikethroughs, overlapping characters, and shadows cover or distort text. Without validating on real documents containing them, performance shifts considerably after go-live.

Third, handwriting has to be handled. A large share of fields on an application form are filled in by hand. Material quoting only printed accuracy tells you nothing about real performance, so ask for handwriting accuracy separately.

Fourth, character positions have to be returned. A recognised string on its own supports neither field extraction nor review. Coordinates arriving with each value are what let a reviewer point at the source on screen.

Fifth, confidence has to be provided at field level. One score for the whole document gives no indication of where to look. Marking low-confidence regions separately is what makes exception-based review possible.

Sixth, it has to sustain the volume. Look at peak-hour concentration rather than the monthly average. Production measurements show 100 pages in 3.5 minutes, 500 pages in 17 minutes, and roughly 1,800 pages per hour.

How OCR is applied in practice

Put preprocessing in front of recognition

The fastest way to raise recognition accuracy is not to change the model but to clean up the input image. A real pipeline converts the document format to images, splits and normalises pages, corrects skew and rotation, removes noise, and only then proceeds to recognition.

Without that front end, no model at the back can recover what was lost. Where fax and mobile photographs make up much of the intake, preprocessing quality decides overall performance.

Design character recognition and structure recognition together

Pulling characters out in sequence scrambles the reading order in tables and multi-column layouts. Real configurations analyse layout first, separating headings, body text, tables, and figure regions and establishing the reading order, then recognise the characters within each region.

Done this way, recognition output flows directly into structured data. It is far more useful for later field extraction and retrieval than a flat list of strings.

Add post-processing rules alongside

Recognition output carries recurring error patterns. Digits confused with letters, thousands separators misread, the same field going wrong repeatedly on a particular form.

Catching these with post-processing rules is faster than retraining the model. Fields with a fixed shape — the digit count of a registration number, the pattern of a date, the separators in an account number — are largely filtered out by rule-based validation.

OCR in the Korean environment

Korean business documents carry several distinctive conditions.

Format first. Korean word processor files are widely used and require separate handling from the conversion-to-image stage onward. Government forms and tax returns include circled line numbers and multi-level table structures.

Then notation. Read a document that uses a period as a thousands separator literally and the entire magnitude of an amount is wrong. Documents printed without the separator in a resident registration number are common. Notation like this becomes a business error directly unless corrected by rule.

Finally the intake path. Fax is still in use, and branch scans and mobile photographs coexist within one workflow. This is an area where a vendor that has trained and validated repeatedly on domestic documents holds an advantage.

Frequently asked questions