Turning a Stack of Scanned Pages into a Searchable PDF

A phone camera and a flatbed scanner solve two different problems, and neither one makes the result searchable by itself.

By PulsarKit · Published September 8, 2026 · Updated October 16, 2026

A photo of a page is not a scan

Held above a desk, a phone camera sees a trapezoid, not a rectangle, and it records the colour of the room along with the page rather than the clean white of the paper. Getting from that photo to something that reads like a real scan means finding the four corners of the sheet in the frame, undoing the perspective distortion so the page comes out square, and then applying a local brightness threshold that turns the paper white and the ink black. Without that, you are left with a grey, slightly warped photograph of a page sitting on a desk. None of it happens by taking a clearer picture. It is a specific correction applied after the shot.

One page at a time, corrected as you go

Automatic corner-finding gets the trapezoid right most of the time. It gets it wrong on patterned tablecloths, dark wood, or a page that runs off the edge of the frame, which is common enough with an oversized book page that it is worth expecting rather than treating as a failure. When it misses, dragging four handles onto the real corners by hand fixes a single page without redoing the rest of the batch. Pages built up this way accumulate in a list that can be reordered, rotated a quarter turn, or retaken in place; a page that still ends up in the wrong spot can be moved later with Organize PDF. Retaking replaces just that one page rather than forcing the whole stack to be scanned again from the start.

A scanned PDF still can't be searched

Once every page is captured, corrected and assembled into one PDF with JPG to PDF, the result is still, technically, a stack of pictures. There is no text layer underneath any of it, so nothing in the document can be selected, copied, or found with a search. That surprises people who assume "scanned to PDF" and "searchable PDF" mean the same thing. They don't, and the gap between them is what OCR closes. Running OCR PDF over the finished scan adds an invisible, precisely positioned layer of recognised text behind the image on each page. The page still looks the same, but now the words on it can be selected, copied, and searched.

Getting recognition right the first time

Recognition accuracy depends heavily on how flat, well-lit and evenly angled the source image is, which is what the scanning and perspective-correction step upstream is meant to guarantee. A properly straightened, cleanly thresholded page recognises far better than a photo taken at an angle in poor light. Deskewing and auto-rotation running automatically before recognition catch most of the remaining tilt. It also pays to pick the right language, or up to three if a document mixes them, since recognition tuned for the wrong alphabet does noticeably worse even on otherwise clean pages.

Choosing skip, redo or force before you commit

A batch of scanned pages is rarely uniform. Some might already carry a text layer from an earlier pass, and re-recognising those wastes time without changing anything useful. Leaving pages that already have text untouched and only recognising the pure-image ones is the safe default for a mixed batch. When an existing text layer is known to be wrong, garbled by a bad earlier OCR pass for instance, forcing recognition to ignore whatever text is already there and redo every page from the image is the option that fixes it. The alternative leaves the bad layer in place underneath a second, unused one.

Common mistakes at the camera

Most bad scans are decided before any software sees them. Shooting under a single lamp puts the shadow of the phone, or of your own hand, across the page, and no threshold can tell that shadow from ink; daylight from a window, or two lights from opposite sides, removes the problem at the source. Glossy paper reflects the light back into the lens as a white patch that erases the text underneath, so tilt the book or move the light until the glare leaves the page. A thick book never lies flat on its own: text near the spine curves and shrinks, and a camera records it as a smear. Pressing the opposite page down with a hand or a weight and photographing one page at a time, instead of a two-page spread, keeps every line straight enough to recognise. Put the book on a plain surface that contrasts with the paper, since patterned cloth and dark wood grain are exactly what fools corner detection. Keep the phone parallel to the page rather than angled, and do not shrink the photos before recognition: it is the thin strokes of the letters that recognition reads, and a downscaled copy loses them first. Finally, shoot in reading order and check the page count at the end; a missing page is far easier to retake now than to discover after the OCR pass.

What happens to the file during recognition

The two halves of this job run in different places. Assembling the pages with JPG to PDF and fixing their order in Organize PDF happen entirely in the browser; the photos are never uploaded. Recognition is different: a multi-page scan needs a real OCR engine, so OCR PDF sends the assembled PDF to our server over an encrypted connection, up to 300 MB and 500 pages per job. The uploaded file is deleted as soon as the job finishes, and the result stays downloadable for one hour and is wiped within about two hours whether you download it or not. The pages come back in the same number and order, each with an invisible text layer; Straighten skewed pages and Auto-rotate pages are on by default and also set crooked or upside-down pages upright. A book shot on a phone reaches the size limit long before the page limit: at about three megabytes a photo, a hundred pages already fill 300 MB. So assemble it in parts of about a hundred photos, or cut a finished PDF with Split PDF, and recognise each part; Merge PDF joins the results in order, up to 500 MB in total. Pick the languages first, up to three for a book that mixes them, and turn on the sidecar text file to get the recognised words as a plain .txt: it is the quickest way to check a sample page before committing the rest.