OCR vs converting a PDF to Word

Two kinds of PDF that look identical

Almost every confusion about OCR comes from one fact: a PDF can store a page in two completely different ways, and on screen the two look the same.

The first kind has a text layer. The file contains real text objects — a string of characters, a font, a size, and coordinates saying where on the page each piece is drawn. The page is a set of instructions for painting letters. A program can read those characters back out exactly, because they were never anything but characters.

The second kind is page images. This is what a scanner, a phone camera or a fax produces: each page is one photograph, wrapped in a PDF. There is no text anywhere in the file. What looks like the letter “e” is a pattern of dark pixels that a human eye recognises as an “e”; to the software it is no different from a picture of a cat.

The two-second test

Try to select a word by dragging across it in any viewer. If individual words highlight, the file has a text layer. If the cursor draws a rectangle instead — a selection box, not a text selection — there is no text layer and you are looking at an image.

Two tools here confirm it without guesswork. Page count reports whether the document has a text layer. PDF to text returns nothing useful for a scan — usually an empty result, sometimes a stray line the scanning software added. That empty result is not a bug; it is the correct answer to “what text is in this file?”

Documents can be mixed. A contract written in a word processor, then signed and scanned, may have a text layer on most pages and none on the scanned one. Check the page you care about.

What OCR does

OCR stands for optical character recognition. It takes an image, finds the regions that look like text, splits them into lines and then into individual glyph shapes, and compares each shape against a model of what characters look like. For every character it emits a best guess and a confidence figure — how sure it is.

The word guess is the important one. OCR does not recover the original text, because the original text is not in the file. It produces a new text that it believes matches the picture — usually very good, occasionally wrong, and nothing in the output tells you which is which unless you inspect the confidence values.

Accuracy depends on things you can mostly see for yourself:

• Scan quality. Resolution, contrast, skew, shadows from a phone camera, speckle from a cheap scanner, and text photocopied a few times too often.
• Language and typeface. An engine has to be told which language to expect; the wrong model produces confident nonsense. Unusual fonts and handwriting are much harder than plain printed prose.
• Layout. This is where OCR most often goes wrong. Reading order has to be inferred from position, so a two-column page can come back with the columns interleaved line by line. Tables are worse: cell boundaries may be ruled lines or nothing but whitespace, and a table often arrives as a stream of numbers with the rows and columns gone.

The standard open-source engine is Tesseract, which is what most free and self-hosted OCR tools use underneath. Whichever engine you use, treat the output as a draft: proofread it, and look especially at digits and at character pairs that look alike in print.

What a PDF to Word conversion does

A PDF to Word converter does something quite different. It reads the text objects that are already in the file, together with their fonts and coordinates, and rebuilds an editable document from them.

The characters are known exactly — they were stored as characters. What has to be worked out is structure, because a PDF page says “draw this string at this position”, not “this is a paragraph” or “this is the second cell of row four”. So a converter infers:

• paragraphs, from lines that share a font and sit at a consistent leading and indent;
• headings and emphasis, from larger or heavier fonts;
• lists, from repeated bullet or number glyphs in a hanging indent;
• tables, from ruled lines or from text that lines up in columns down the page.

That inference is also a guess, but a different one from OCR’s. A converter can be wrong about where a paragraph ends; it is not wrong about which letters are in it.

Run a converter on a scan and it fails for a simple reason: there is nothing to convert. It does not read pixels, so it finds no text and produces either an empty document or a Word file with one full-page picture per page — no more editable than the PDF was.

Which one you need

The decision path is short:

1. Try to select a word in the PDF.
2. If words highlight, you need a conversion. The text is there; you want it in an editable format.
3. If you get a selection box, you need OCR first. Conversion has nothing to work with until OCR has created a text layer.
4. If you only want the words — to quote, search or paste them — PDF to text is enough; you do not need a Word file at all.

If you are not sure how a file was produced, the document’s own metadata often says. The producer field frequently names the scanner or the application that wrote the PDF — see what PDF metadata contains.

What iBuildPDF can and cannot do here

Plainly: iBuildPDF cannot do either half of this job for you.

There is no OCR tool on this site. None of the tools recognises characters in an image. PDF to text extracts a text layer that already exists; it cannot create one. On a scanned PDF it returns nothing, and no setting changes that.

The PDF to Word converter is not available. It is listed as coming soon and is switched off, as are the PDF to Excel and PDF to PowerPoint entries. Do not plan work around them.

What to do instead, in the order worth trying:

• Ask the sender for the original. If someone scanned a document they wrote, the source file still exists, and it will be more accurate than anything OCR or conversion can reconstruct.
• Use a dedicated OCR tool. Tesseract is free and widely packaged; commercial products generally handle tables and poor scans better. Choose one that lets you set the language.
• Improve the scan first. Re-scanning at a higher DPI, flat and square, with good contrast, does more for accuracy than any post-processing.
• Check with page count before you start, so you know which problem you actually have.