OCR, short for Optical Character Recognition, is a technology that converts text from images, scans, photographs, screenshots or PDF files into machine-readable text. Instead of forcing a person to retype everything manually, OCR tries to recognise letters, numbers and other characters directly from the visual input and turn them into text that can be copied, searched, stored or processed further.

At first glance, OCR can seem like a simple feature. You upload an image with text and the system turns it into something editable. In reality, OCR solves a much more difficult problem. It has to recognise characters across different fonts, sizes, layouts, languages, image quality levels and often in documents that are not perfectly clean, flat or sharp.

What OCR actually does in practice

OCR is used wherever text exists in visual form but people or systems need to work with it as ordinary digital text. A scanned document is the most obvious example. You can see the text on screen, but without OCR the computer often treats it as just an image. That means it cannot reliably search inside it, copy specific lines or automatically extract useful fields.

OCR addresses this by analysing the image and trying to convert visible characters into a text layer. That makes it possible to pull text from a scanned contract, read an invoice number from a document, capture the total from a receipt or copy a serial number from a photographed product label.

This is why OCR is not only about digitising paper archives. It also matters in administration, accounting, logistics, e-commerce, banking, customer support, public administration, legal work, healthcare and technical documentation.

Why OCR is not the same as scanning

Scanning and OCR are often confused, but they are not the same thing. Scanning creates an image of a document. OCR tries to extract text from that image.

If you scan a paper contract without OCR, the result may look like a proper digital document, but technically it still behaves more like a picture. A person can read it, but the system may not know what text it contains. Once OCR is applied, the file can gain a text layer. That allows searching, copying and further processing.

Practically speaking, scanning creates a digital copy of the paper. OCR tries to understand what text is written on that copy. That is why a PDF may look perfectly readable but still fail at text search or copy-paste if OCR has not been applied.

How OCR works

OCR usually happens in several steps. The exact process depends on the tool, the document quality and the underlying technology, but the general principle is similar.

1. Capturing the visual input

The system first needs an input image. That may be a scan, a photo, a screenshot, a PDF, a receipt, a label or a book page. Input quality matters a great deal. The clearer, straighter and sharper the source image is, the better the chance of correct recognition.

2. Image preprocessing

Before text recognition begins, the system often cleans or adjusts the image. It may increase contrast, reduce noise, straighten a tilted page, separate text from the background or detect where text blocks actually sit on the page.

This step matters especially for phone photos, older scans, receipts, forms or any material that is not perfectly clean and flat.

3. Recognising characters and words

The OCR engine then analyses sections of the image and tries to identify letters, numbers, punctuation and other characters. Older OCR systems relied much more on pattern matching against expected character shapes. Modern OCR systems increasingly use machine learning and neural networks, which cope better with varying fonts, noisy input, lower-quality images and more difficult page layouts. :contentReference[oaicite:1]{index=1}

4. Reconstructing reading order and structure

Recognising characters alone is not enough. OCR also needs to determine the order of words, lines, paragraphs and sometimes columns. With a simple block of text, that is relatively straightforward. With invoices, forms, tables, newspapers or multi-column PDF files, it becomes much harder.

5. Producing the output

The final output may be plain text, an editable document, a searchable PDF with a text layer, a spreadsheet, JSON, XML or another structured format. The output depends on what the recognised text will be used for.

Where OCR is used most often

OCR has a very wide range of uses, including cases where people do not even realise it is being used behind the scenes.

• Document digitisation – converting paper archives into searchable electronic files.
• Accounting – extracting data from invoices, receipts, purchase orders and delivery notes.
• Banking – reading documents, statements, forms and applications.
• Logistics – recognising labels, shipment numbers and delivery documents.
• E-commerce – processing product labels, returns documents, invoices and supplier materials.
• Legal services – turning contracts and case files into searchable text.
• Healthcare – digitising reports, forms and archival records.
• Education and libraries – digitising books, manuals, articles and historical material.
• Mobile apps – reading text from photos of receipts, business cards, documents and labels.

OCR for invoices and receipts

One of the most common practical uses of OCR is invoice and receipt processing. A system may recognise supplier details, tax ID, issue date, due date, invoice number, payment reference, total amount or line items.

But it is important to distinguish between plain OCR and actual data extraction. OCR reads the text. A more advanced system still has to understand what the recognised text means in the document context.

Example:

• OCR output: “Due date 15 May 2026, total amount EUR 1,250.”
• Data extraction: the system identifies that 15 May 2026 is the due date and EUR 1,250 is the payable amount.

That is a major difference. Reading the text is not yet the same as understanding which field belongs where in an accounting or ERP workflow.

OCR and PDF documents

A PDF can be technically very different depending on how it was created. Some PDFs already contain real text. In that case, the text can be selected, copied and searched without OCR.

Other PDFs are just images of pages, typically scanned paper documents. They may look like ordinary PDFs, but the text is not stored as text. In that case OCR is needed.

There is also a hybrid case: a PDF that shows the original scanned image but includes an invisible text layer generated by OCR. The user sees the scan, but search and copy-paste still work because the OCR layer sits behind it.

OCR and handwritten text

Handwritten text is more difficult than printed text. Different people write differently, handwriting can be irregular, slanted, abbreviated or simply hard to read. That is why handwritten recognition is often discussed separately under the term HTR, meaning Handwritten Text Recognition. Microsoft’s OCR and document tools explicitly support both printed and handwritten text extraction, but accuracy still depends heavily on input quality and writing style. :contentReference[oaicite:2]{index=2}

Modern systems are much better at handwriting than older OCR tools were, but printed text is still generally the easier case. With handwriting, poor photos, old documents, weak contrast or complex forms, the chance of error rises significantly.

OCR and tables

Tables are difficult for OCR. It is not enough to read the text inside cells. The system also has to reconstruct the table structure – rows, columns, headers, merged cells and relationships between values.

This matters in invoices, price lists, technical sheets, bank statements or measurement reports. If OCR reads the cell text but loses the structure, the result may be far less useful for further processing.

That is why table-heavy workflows often use more advanced document AI or document intelligence tools rather than basic OCR alone. These systems try not only to read text, but also to understand document layout and structure. :contentReference[oaicite:3]{index=3}

OCR and barcodes

OCR is sometimes confused with barcode or QR code recognition, but they are not the same thing. OCR recognises letters and numbers from the visual shape of text. A barcode or QR code is a special machine-readable symbol that is decoded differently.

In practice, though, these technologies often work together. In logistics, for example, a system may decode a barcode and at the same time use OCR to read nearby text such as product name, batch number or address.

OCR and artificial intelligence

OCR has gradually moved beyond basic character recognition toward more advanced systems that rely on machine learning. Modern tools often go further than simply asking “What text is written on this page?” They also try to answer “What does this text mean in the context of this document?” Google and Microsoft both position newer OCR-related services as part of broader document understanding pipelines, not just plain text extraction. :contentReference[oaicite:4]{index=4}

That matters especially in more complex business workflows. On an invoice, for example, it is not enough to read every word. The system should also recognise which number is the invoice number, which is the due date, which amount is without VAT and which amount is actually payable.

This is where OCR connects with document processing, computer vision, machine learning and language-oriented AI systems.

OCR versus multimodal models

OCR and multimodal models are related, but they are not the same thing.

OCR focuses mainly on turning text from images into machine-readable text. A multimodal model can then work with that text further, interpret it, combine it with the image itself, answer questions about the document or explain what the document means.

Example:

• OCR: reads the text from an invoice.
• Multimodal model: can summarise the invoice, point out a missing field or answer whether it is overdue.

OCR is therefore often one technical layer that enables further document processing. By itself, it does not necessarily mean the system truly understands the document.

Why OCR sometimes makes mistakes

OCR is not error-free. Mistakes happen especially when the visual input is poor or the document structure is complicated.

Typical causes include:

• blurred or low-resolution images,
• skewed pages,
• bad lighting or low contrast,
• complex layouts,
• mixed languages,
• unusual fonts,
• damaged, noisy or old source material,
• dense tables and forms.

That is why OCR output should not automatically be treated as perfect just because it looks plausible. In high-stakes cases, verification still matters.

What OCR cannot guarantee on its own

OCR is a powerful enabling technology, but on its own it does not guarantee that the final result will be correct, complete or properly understood.

OCR by itself does not automatically:

• verify whether the recognised text is factually correct,
• understand legal priority between documents,
• know which date is the most important one in context,
• decide which extracted value belongs in which business field,
• replace human review in sensitive legal, financial or medical situations.

That usually requires additional layers such as field extraction, validation rules, metadata, document intelligence, business logic or human oversight.

Why OCR matters beyond technical roles

OCR is one of those technologies many people only notice when something goes wrong. Yet it plays an important role in everyday business operations, archives, search, compliance and automation.

Understanding OCR helps explain why one PDF can be searchable while another is not, why copied text from a scan may be broken, why invoices sometimes need verification after automated extraction, or why a photographed document can still be difficult for systems to process reliably.

That matters not only to developers, but also to operations teams, finance staff, legal departments, support teams, content managers and anyone working with documents at scale.

Related terms

• Scanning – creates a digital image of a document. It is closely related to OCR because scanning often comes first, while OCR adds machine-readable text on top of the scan.
• HTR – short for Handwritten Text Recognition. It matters because handwriting is usually harder to process than printed text and is often treated as a more difficult neighbouring problem.
• Document AI – broader document-processing systems that do more than read text. They often combine OCR with layout analysis, field extraction and structured understanding.
• Computer vision – the wider field focused on analysing images and visual content. OCR belongs here because it starts from visual input rather than from ready-made digital text.
• Machine learning – modern OCR increasingly relies on machine learning models instead of only rule-based character matching. For a wider foundation, see Machine learning.
• PDF – important in OCR workflows because some PDFs already contain real text while others are only scanned page images and require OCR to become searchable.
• Text extraction – a related goal, but not always the same thing. OCR extracts text from images, while text extraction in a broader sense may also refer to pulling text from digital sources that already contain real text.
• Multimodal model – relevant because multimodal AI can build on OCR output and then interpret or explain the recognised document content further.