Fast, Advanced or Compatibility: picking an OCR mode

A 40-page scanned tender document, run on a phone during the train journey home, came back with half the pages readable and half looking like Hll. lL,cn te. The same file on a desktop that evening read almost perfectly, first try, same settings. Nothing changed about the PDF.

That gap is not random. Advanced OCR offers three recognition modes and picks its working resolution from the machine it finds itself on, and those two facts explain most of the difference between a clean read and a page of nonsense.

What the three modes actually do differently

Fast runs one recognition pass. Before that pass, the page is prepared automatically — orientation is detected, skew is straightened, contrast is lifted off the background. That preparation is not optional and not part of the mode choice; it happens every time. Fast simply accepts the first result it gets.

Advanced is a multi-pass mode. Instead of one prepared version of the page, it builds several using different cleanup recipes and keeps whichever recognised best. The recipes are specific:

• grayscale — colour thrown away, nothing else touched
• high-contrast — grayscale plus an automatic contrast stretch
• adaptive-threshold — grayscale, contrast, then a local black-or-white decision made in small windows across the page rather than one threshold for the whole sheet
• sharpened — grayscale, contrast, then edge sharpening
• shadow-corrected — the background lighting gradient estimated and subtracted, then contrast

It does not blindly run all of them on every page. A page that already read cleanly is left alone. A second attempt is triggered when the mean word confidence falls below 75, when the output is full of punctuation noise rather than letters, or when a page that clearly is not blank returns almost no text. On a long document that rule is the difference between a few minutes and most of an hour.

Compatibility goes the other way. It renders at a lower resolution and works strictly one page at a time. You lose accuracy. You gain a tab that does not die at page 60.

Why Advanced is sometimes greyed out

If the Advanced option refuses to select and says the device has too little memory for multi-pass recognition, that is a hardware reading, not a paywall. Recognition happens inside your browser, so the constraint is your own RAM and cores, and the tool measures them before offering you anything.

That measurement also sets three other numbers you will run into:

The last column quietly decides your accuracy. An A4 page scanned at 300 DPI is roughly 3500 pixels down its long edge. On an older phone it gets scaled to 1600 before a single character is looked at — less than half. Small print, footnote markers and the digits inside a table are exactly what disappears at that scale. Same PDF, same mode, worse result, and the reason is arithmetic rather than anything you did wrong.

Choosing without trial and error

That last row is the habit worth building. Fast on one page tells you the page confidence in seconds. If it comes back at 92% with two words flagged, Advanced has nothing left to add and will cost you five times the wait. If it comes back at 58%, escalate.

The language picker is not a select-all box

Up to three languages can be active at once. Each one downloads its own recognition data the first time you use it, and each one widens the set of shapes the engine has to weigh on every single word.

Ticking English, Hindi and Marathi for a document that is entirely in English does not make it safer. It makes recognition slower and gives the engine three ways to misread the same letter. Pick what is genuinely on the page. A tehsil letter with an English letterhead over a Hindi body needs both; an export invoice in English needs one.

The first run of any language is the slow one, because the data has to be fetched into your browser before it can be used. Your browser then caches it. That wait is the price of the document never being uploaded — a server-side OCR service starts instantly only because your file is already sitting on its hardware.

What no mode will rescue

Two honest limits, because switching modes in circles is a good way to lose an evening.

Handwriting is not what this engine does. It is built for printed characters, and results on cursive or hurried block capitals are unreliable in a way that is worse than useless — wrong output that looks confident is much harder to catch than no output at all.

A source below roughly 200 DPI cannot be recovered by any recipe either. The information is not in the file. Re-scanning at 300 DPI beats every setting on the page, and if a photograph is your only option, fill the frame with the sheet, lay it flat, and use even light rather than one overhead bulb throwing a shadow across the middle.

There is a third check worth making before you run anything. If you can select text in the PDF with your cursor, it already carries a text layer and OCR is the wrong tool entirely — PDF Text Extract will pull that text out exactly, with no recognition guesswork involved. Advanced OCR is upfront when this happens: a page read from the document's own text layer is labelled as such rather than being passed off as recognition.

When a page comes out sideways

Orientation detection needs a reasonable amount of text before it can decide which way up a page is. A near-empty page — a divider, a signature sheet, a cover with six words on it — can defeat it, and recognition then runs on a page rotated 90 degrees.

For a single image, straighten it first with Image Rotate and feed the corrected file back in. For a scanned PDF where several pages went in sideways, the server-backed PDF OCR route is often the easier path, since it takes the whole document with one language selection and hands back a searchable PDF or an editable Word file rather than block-by-block text you have to reassemble.

The rest of the scanning and recognition tools sit together on the scanner tools page if you need to line one up behind another.

A workable order of operations

1. Check whether the text is already selectable. If it is, you do not need OCR at all.
2. Run Fast with the correct language, or languages, ticked.
3. Read the page confidence and the count of words flagged for checking.
4. If a page reads poorly, re-run in Advanced rather than accepting it.
5. If the tab struggles or the run stops partway, drop to Compatibility and let it take its time.
6. Fix the flagged blocks in the editable boxes before you export anything.

Mode choice is worth maybe fifteen per cent of the outcome. Scan quality is worth the rest.