Optical Character Recognition
Optical Character Recognition, commonly abbreviated OCR, is the automatic conversion of images containing writing into editable text. A photo of an invoice thereby becomes characters that a computer can search, copy, and process further.
A photo of a sheet of paper is at first, for a computer, nothing more than a collection of colored pixels. It sees light and dark patches, but no letters and no words. Optical Character Recognition, usually abbreviated OCR, closes exactly this gap. The process recognizes the written characters within an image and outputs them as actual text. After that, the content can be searched, copied, translated, or sorted into a table. The German technical term for this is optische Zeichenerkennung.
Why paper is still a data problem
To this day, a great deal of information is still created on paper or as a photo of it. Invoices, contracts, medical records, forms, receipts, and newspaper archives all belong to this category. Without text recognition, a human would have to type all of this out by hand. That is expensive, slow, and prone to error. OCR turns this dormant stock of material into searchable data.
This is particularly important in the financial world. Banks and insurance companies review millions of documents every day, often scanned or photographed with a phone. Only once amounts, dates, and names exist as text can software automatically post them or check them for irregularities. Studies estimate that manually capturing an invoice costs several euros. With automated processing, this amount drops to mere cents.
A second reason is accessibility. Blind people use programs that read screen contents aloud. A scanned book without text recognition remains silent for such programs. OCR makes it readable.
From pixel to letter
Classic systems work in several steps. First, the image is prepared: it is straightened, brightened, and cleared of stains. Then the software searches for the areas where text actually appears. Next, it breaks lines down into individual characters and compares each one against stored shapes. This approach works well for clean print in a known typeface.
Modern OCR instead uses neural networks. These are learning programs that figure out for themselves, from millions of example images, what an A looks like. They do not look at individual characters in isolation, but at entire lines in context. This allows them to cope with handwriting, tilted phone photos, and faded paper as well. If the model only recognizes an unclear shape, context helps: after “Inv”, “oice” is more likely than “oicee”.
A common misconception is that OCR understands the text. It does not. It only delivers the sequence of characters, not their meaning. To know which number on an invoice is the total amount, a further step is needed. This area is called document understanding, and today language models are used for it.
OCR in the phone, in the archive, and in bookkeeping
The technology is built into devices that almost everyone uses. If you point a smartphone camera at a street sign, the text can be tapped and translated. Translation apps work the same way. Even searching the photo gallery for a word only works because text recognition has run over the images in the background.
In everyday life, OCR also shows up at the parking garage barrier that reads the license plate, and when photographing a bank transfer slip in a banking app. Libraries use the process to digitize old runs of newspapers. Government agencies use it to further process submitted forms.
In tech news, the term usually appears in connection with automation. Vendors advertise that their software posts invoices without human involvement. What matters here is always the error rate. Ninety-nine percent correctly recognized characters sounds good, but for a twenty-digit account number it still means regular errors. That is why serious systems check critical fields against other data sources or present them to a human for review.