OCR PDF — Make Scanned PDFs Searchable

What the engine is doing while it reads a page

Recognition runs in stages, and knowing them explains most of the results people get.

First the page is reduced to black and white. A greyscale scan has to be resolved into ink and paper, and a threshold is chosen for the page — which is why a shot with a shadow across it can lose a whole corner to the paper side of the line while the lit half reads perfectly.

Then the layout is analysed: blocks of text are separated from pictures and rules, blocks are broken into lines, and lines into candidate words and characters. This stage decides reading order, so a two-column page misjudged here produces text that is individually correct and collectively scrambled.

Only then is each character shape classified, and the classifier is not working alone. Its raw guesses are weighed against a model of which letter sequences are plausible in the language, which is how a smudged word is usually recovered from its neighbours.

Every stage depends on the one before it. A page that fails at binarisation cannot be rescued by good classification.

Why 300 DPI is the number everyone repeats

The threshold sounds arbitrary until you convert it into pixels per letter. Type is measured in points at 72 to the inch, and ordinary body text is around ten points, so a capital letter occupies roughly a tenth of an inch of page height.

Scan at 300 dots per inch and that letter is about thirty pixels tall — enough for the classifier to see the gap in an "e", the tail on a "y", the difference between "rn" and "m".

Scan at 150 and it is fifteen pixels. The shapes still look like letters to a person, because a person is reading whole words in context, but the fine features the classifier depends on are now one or two pixels wide, and one or two pixels is what a slightly soft focus removes entirely.

Below that, adjacent strokes merge into single blobs and the output becomes unusable. This is why rescanning a poor original beats every adjustment available afterwards: the information the engine needs was never captured, and nothing downstream can invent it.

A worked example: a box of archived correspondence

Forty pages of typed letters from a filing cabinet, scanned on an office multifunction device. At 300 DPI in greyscale, forty pages of A4 lands somewhere in the tens of megabytes — comfortably within the 100 MB ceiling, though colour scanning of black-ink documents can push a batch this size close to it for no benefit whatsoever.

Run OCR and the file comes back looking identical. To confirm it worked, pick a word you can see on a middle page — a surname, a place, a reference number — and search for it. If the reader jumps to it, the text layer is there and positioned correctly.

Then test the weakest page rather than the best one. Find the letter with the faintest print or the worst skew, copy a paragraph out of it, and paste it somewhere you can read it plainly. That page sets the accuracy you can actually rely on across the batch, and it tells you whether the archive is genuinely searchable or merely searchable in the places that were already easy.

Order of operations around OCR

OCR should see the sharpest version of the page that exists, which puts it early in any sequence.

Run it before Compress PDF, never after. Compression downsamples images, and downsampling is precisely the operation that destroys the fine stroke detail the classifier needs. Recognise first, then compress — the text layer survives compression intact, because it is text rather than picture, so you keep the searchability and still get the smaller file.

Straighten the pages first as well. A line of text running at a slight angle crosses several rows of pixels, and line segmentation has to cut somewhere; Rotate PDF handles pages fed in the wrong orientation, which is the common case with a batch scanned in one pass.

If the source is a set of photographs rather than a scan, assemble them with JPG to PDF at full resolution and resist the urge to shrink them for convenience along the way. The convenience is paid for in accuracy.