What is Optical Character Recognition?

Optical character recognition (OCR) is the conversion of images of text, such as scans and photographs, into machine-readable character data.

OCR answers one question: which characters appear on this page. It does not decide what the document is, which values matter, or whether any of them are correct.

What OCR actually produces

A modern OCR engine returns more than a flat string. Typical output includes the recognized characters, their bounding coordinates on the page, a per-character or per-word confidence value, and some description of structure such as lines, blocks and reading order.

That structural output is why OCR quality is not a single number. An engine can read every character correctly and still return them in an order that makes the result useless, which happens routinely with multi-column layouts, sidebars and forms where labels and values are visually adjacent but far apart in the underlying stream.

Layout analysis and reading order

Before recognition, the page is segmented: text blocks separated from images, columns identified, tables detected, headers and footers set aside. This is layout analysis, and it determines whether the characters come back in a sequence a human would recognize as the document’s content.

Tables are the standard difficulty. A borderless table is visually obvious and structurally invisible, so cell boundaries have to be inferred from whitespace alignment. Merged cells, wrapped text within a cell and continuation across a page break all break naive detection.

Language and script detection

Multilingual documents need the language identified before or during recognition, because character models are language-specific and so are the dictionaries used to correct ambiguous shapes. Mixed-script documents, common in Arabic and Gulf-region business paperwork where English product names sit inside Arabic prose, need per-region detection rather than a single document-level language setting.

Right-to-left scripts add reading-order complexity that is separate from character accuracy. A system can identify every Arabic character correctly and still emit them in an order that reverses meaning when interleaved with Latin text and digits.

Why OCR output is not structured data

This is the distinction that matters commercially, and it is the one most often blurred. OCR gives you text. A process needs values: this invoice number, that total, these line items, this signature date.

Getting from one to the other requires deciding what the document is, which is document classification, and then locating and interpreting specific values, which is document data extraction. Together with validation and routing, that pipeline is intelligent document processing. OCR is its first substantive stage, not a substitute for it.

A useful test: if the output of your system is searchable but a person still has to read each document to key values into another system, you have deployed OCR, not document automation.

Accuracy depends more on input than on engine

Teams comparing OCR engines usually find the differences smaller than expected, because image quality dominates. Resolution below roughly 200 dots per inch, heavy JPEG compression, skew, shadows from phone photographs, faint thermal-printer output and scanning at the wrong contrast all cost more accuracy than the gap between competent engines.

This is why preprocessing carries disproportionate weight. Deskewing, despeckling, contrast normalization, and rejecting images that fall below a quality floor before recognition is attempted, tend to improve end-to-end results more than switching engines.

The corollary is worth stating plainly: a project that fixes its capture channel, for example by replacing photographed documents with a scanning app that enforces resolution and framing, often gains more than one that spends the same effort on model selection.

What commonly goes wrong

Accuracy is quoted at character level. Ninety-nine percent character accuracy sounds excellent and can still mean a meaningful proportion of documents contain at least one wrong field, because errors concentrate in the low-quality minority rather than distributing evenly.

Confidence is treated as correctness. A high OCR confidence value indicates the engine’s certainty about character shapes. It says nothing about whether the value is the right one for the field. See confidence score.

Handwriting is assumed to be in scope. Recognition of handwritten text is a different problem with materially different accuracy, and constrained fields such as boxed dates behave very differently from free-form annotation.

Contellect One performs recognition in the read stage of intelligent document processing.