A scanner gives you a picture. To software, that picture is a grid of pixels with no notion that some of them form the word invoice. OCR services in the UAE exist to close that gap: the recognition step reads the shapes, decides which characters they are, and writes the result out as text the rest of your systems can use.
What OCR actually does
The engine segments the page into blocks, then lines, then individual characters. Each shape is matched against a trained model of the script and scored. Nothing you can see on the page changes. What changes is that the document becomes searchable, quotable, and readable by any system that expects text rather than an image.
That is the whole job. OCR does not decide what kind of document it is looking at, does not know which number on an invoice is the total, and does not file anything. Those are separate steps that sit on top of recognition.
What drives the quality of an OCR result
Recognition quality is less a property of the software than of the paper in front of it. Two organisations running identical engines on identical settings can get very different results, and the difference is almost always in the originals. The factors that matter most:
- Contrast between ink and paper. Faded toner, tea-coloured archive stock and third-generation photocopies all narrow that gap until characters start to break up.
- Capture resolution. Office text is normally captured at 300 dpi; small print, footnotes and Arabic diacritics benefit from more detail than a human reader strictly needs.
- Typeface. Clean serif and sans-serif print reads well. Dot-matrix output, condensed legal print, decorative headings and fax receipts are considerably harder.
- Page geometry. Skew, curl at a tight binding and sheets fed at an angle distort character shapes before the engine ever sees them.
- Background interference. Cheque security patterns, letterhead watermarks, ruled lines running through text and bleed-through from the reverse side all read as noise.
- Layout. Multi-column pages, forms and tables need zoning, or the text comes back snaking across columns in an order nobody wrote it in.
- Language mix. A line that switches from Arabic to English and back is harder than either script on its own.
Where OCR struggles, plainly
Anyone who tells you recognition reads everything has not looked inside your boxes. It is more useful to know in advance which of your material will sail through and which will need a person, because that is what the quotation is really made of.
| Source material | What OCR typically returns | How the project handles it |
|---|---|---|
| Clean laser-printed office documents, Arabic or English | A complete text layer with isolated flagged characters | Automated pass with sampled quality checks |
| Faded carbon copies, thermal receipts, third-generation photocopies | Dropped and broken characters, whole words missing | Rescan at adjusted threshold and contrast; nominated fields typed by an operator |
| Handwritten notes and margin annotations | Little or nothing usable | Routed to ICR or to manual capture, depending on how the writing is formed |
| Stamps, seals or signatures printed across body text | Covered characters are misread or invented outright | Region flagged; an operator confirms the wording underneath |
| Skewed, curled or tightly bound pages | Lost lines at the gutter, merged characters | Deskew and dewarp, or a rescan on a book cradle |
| Forms with ruled boxes and shaded fields | Rules read as characters, values pulled into the wrong field | Form-line removal and zonal templates before recognition |
| Dot-matrix, fax and microfilm-sourced images | Fragmented character shapes and frequent substitutions | Image enhancement first, verification on anything scored low |
Arabic and English records
Arabic is cursive by default. Letters change shape according to their position in a word, ligatures fuse characters that a Latin engine would treat separately, and short vowel marks may or may not be present. Recognition therefore depends on segmenting the word correctly before a single character is identified. Add right-to-left flow carrying left-to-right numerals in the same line, and the layout analysis has to be right before the reading even starts.
UAE records are rarely monolingual anyway. A tenancy contract, a trade licence or a ministry form usually carries an Arabic label above an English value, or two parallel columns saying the same thing twice. We set the recognition language per zone rather than per file, and we treat bilingual output as something to check rather than something to assume.
Handwritten Arabic on a faded carbon copy is the hardest material in this industry, and we would rather say so at the quotation stage than hand back a text layer nobody can rely on. Those pages go to human verification.
What you receive
- Searchable PDF: the original page image with an invisible text layer sitting behind it, so the document still looks exactly like the paper.
- PDF/A where the records are going into long-term archive rather than daily use.
- Plain text, XML or JSON sidecar files when the text is feeding a system rather than a reader.
- A per-page confidence report, so pages the engine found difficult are visible to you instead of quietly averaged away.
- The original images, retained unaltered, so recognition can be re-run later on better software without touching the paper again.
What an OCR project costs to run
Volume, condition, language mix and the depth of verification you want. A clean, single-language, printed archive is among the least expensive work we do. A mixed box of stamped, handwritten and bilingual paper is not, because the cost sits in the human passes rather than the machine one. Most projects are quoted after a sample batch and a page-count assessment, which is also the point at which we can tell you honestly how much of your archive will need a person.
Frequently asked questions
What is the difference between scanning and OCR?
Scanning produces an image of the page. OCR reads that image and produces text. A scanned PDF without OCR looks identical on screen but cannot be searched, copied or indexed by content. Most digitisation projects need both: capture to create the image, recognition to make it useful.
Does OCR work on Arabic documents?
Yes, for printed Arabic. Cursive joining, positional letter forms and diacritics make it harder than printed English, and mixed Arabic-English lines need the recognition language set per zone. Handwritten Arabic is a different problem and generally needs human verification rather than an automated pass.
Can OCR read handwriting?
Standard OCR is trained on printed characters and will usually return nothing usable from handwriting. Neat, boxed handprint can be read by ICR, a related technology built for that purpose. Free-flowing cursive, particularly on poor originals, is normally captured by an operator rather than a machine.
How accurate is OCR?
Accuracy depends almost entirely on the source, not the software. Clean printed originals at adequate resolution read very well; faded, stamped, skewed or handwritten material does not. Rather than quote a single figure, we run a sample batch from your own archive and report what your material actually produces.
Can you OCR documents we have already scanned?
Yes. If you hold image-only PDFs or TIFF files, recognition can be applied retrospectively without touching the paper again. Results depend on how those images were captured. Low-resolution, heavily compressed or bitonal scans of faint originals limit what any engine can recover from them.
Does OCR change how the document looks?
No. In a searchable PDF the recognised text is written as an invisible layer positioned behind the page image. What you see is still the original scan, including stamps, signatures and letterhead. Only the search, copy and index behaviour changes, which matters where the image is the legal record.
Related reading
- ICR and intelligent data captureFor handprinted entries and form fields that standard OCR cannot read.
- Searchable PDF and PDF/A conversionWhat the recognised text layer looks like inside the file you receive.
- Document indexing and metadataFull-text search finds words; index fields find the right document.
- Capture and recognition technologies we use
- Athena Global Technologies, our parent companyCorporate background, leadership and the wider enterprise service range.