Files and output
Start with what the file contains. A PDF can hold images, encoded text or both; its filename does not tell you whether its words can be selected.
OCR — optical character recognition
The process of identifying visible characters in an image and representing them as encoded text. It lets you reuse words from a scan or photo; the recognized result may still need correction.
Scanned PDF
A PDF whose pages include scanned or photographed document images. Some scanned PDFs also have an OCR text layer, while others contain only images.
Born-digital PDF
A PDF created directly from digital content, such as a document editor, rather than by scanning paper. It often contains actual text, although selection and extraction can still be affected by how it was produced.
Searchable PDF
A PDF containing text that a reader can search. A scanned page can become searchable when recognized text is added alongside its image. Searchability does not by itself guarantee accuracy or accessibility.
Text layer
The encoded text stored in a PDF. On a scanned page it may be invisible and positioned over or under the image so selection and search work while the original appearance remains visible.
Page structure
A recognizer must do more than identify letters. It must also decide which marks belong together and how the reader should move through them.
Layout analysis
Identifying regions such as paragraphs, headings, images and tables. Recognizing a table's words and preserving its row-and-column structure are related but distinct tasks.
Reading order
The intended sequence of text. A two-column article should normally follow one column before continuing in the next, rather than alternating across both. Correct words in the wrong order can produce an unusable transcript.
Segmentation
Dividing a page into smaller units such as text regions, lines or characters. Incorrect boundaries can join neighboring words, split a character or mix unrelated parts of a page.
RTL — right-to-left text
Text in scripts such as Arabic or Hebrew that generally flows from right to left. Numbers and embedded Latin text can follow different directions. Display order and stored character order are not necessarily the same.
Image preparation
Preparation changes how the existing marks are presented to recognition. It cannot establish text that the source image no longer contains.
PPI and DPI
Pixels per inch describes image sampling at a physical size; dots per inch describes printed dot density. Scanning settings often use DPI informally. Changing the stored value alone does not add detail to an existing image.
Deskew
Correcting a small tilt so text lines become more horizontal. It is different from rotating an upside-down page or correcting the perspective of a photograph taken at an angle.
Binarization
Converting an image into black and white pixels using a threshold or other decision rule. It may simplify a clear page, but can also lose pale marks or merge nearby strokes.
Preprocessing
Image adjustments made before recognition, such as orientation correction, cropping or contrast changes. A useful adjustment improves the recognizable marks without removing punctuation, accents or other meaningful detail.
Recognition and review
An evaluation describes a particular task and dataset. Read the definition of a score before comparing it with another result or using it to judge your own documents.
Handwriting recognition
Recognition of handwritten characters or words. Connected letters, personal abbreviations and writer-specific shapes make it a different challenge from clear printed text.
Confidence score
An estimate a recognition system may attach to a prediction. Its meaning depends on the system and calibration. It is not automatically the probability that an entire document is correct.
CER — character error rate
A comparison of recognized text with a checked reference, based on character insertions, deletions and substitutions relative to the reference length. Lower is better. A low CER can still hide an important error in a name or number.
WER — word error rate
A similar comparison at the word level. One incorrect character can make a whole word count as wrong. WER and CER therefore describe different aspects of a transcript and should not be treated as interchangeable.
Ground truth
The checked reference used to evaluate recognition. Its transcription conventions matter: how it handles punctuation, spelling, spaces and unreadable text affects the comparison. A flawed reference can produce a misleading score.
Questions and answers
Is OCR the same as translation?
No. OCR reads visible text into an editable form. Translation expresses text in another language. Reviewing recognition first helps avoid translating a misread source.
Is a Word download a copy of the original layout?
Not necessarily. An editable document can contain recognized text without preserving the source's tables, spacing or page design. Check the output against the original when layout matters.
Why can two accuracy scores disagree?
They may use different documents, reference conventions or measurements. Character errors, word errors, omitted regions and reading-order problems are not the same thing.