OCR — Image & PDF to Text

The gap OCR fills

A scanned document is a photograph wrapped in a PDF. To a computer it is pixels: there are no characters in the file, which is why you cannot select text in it, why searching finds nothing, and why every text-extraction tool returns an empty result.

Optical character recognition closes that gap by examining the image and identifying letter shapes. It is a genuinely different operation from extraction — extraction reads characters that are already there, OCR guesses at characters from pictures — and that difference is why OCR is never perfect while extraction is exact.

The documents that need it are almost always the sensitive ones: contracts that arrived as scans, bank statements from an older system, receipts photographed for an expense claim, identity documents for an application, medical letters, historic records. Every mainstream free OCR service handles these by uploading them.

Why the language setting matters most

OCR does not simply match letter shapes. It uses a language model to resolve ambiguity — deciding whether a mark is an `l`, a `1` or an `I` by asking which produces a plausible word. Set the wrong language and that model works against you, turning correct shapes into wrong letters.

Choosing the right language typically moves accuracy by more than any resolution change. If your document is bilingual, run it twice and take the better result for each section; Tesseract can combine languages, but the combination is usually worse than either alone.

Each language pack is roughly 10–15 MB and downloads once, then stays in your browser cache. The first run in a new language is therefore slower than every one after it.

Getting a good result

Input quality dominates everything. A flat, straight, evenly lit scan at 300 DPI in greyscale reads dramatically better than a phone photo taken at an angle in a dim room, and no setting compensates for a bad source.

Straighten the page before scanning — even a few degrees of skew hurts, because the engine looks for horizontal text lines. Crop away everything that is not the page. Avoid shadows across the text, which is the single most common problem with photographed documents.

Resolution has a sweet spot. Below about 150 DPI, small text loses the detail needed to distinguish similar letters. Above 300 DPI you mostly add processing time. For a page of ordinary body text, 200 to 300 DPI is right.

The confidence figure reported after each run is worth reading. Below about 70% the output will need real editing; above 90% it is usually close to clean.

What OCR will still get wrong

Handwriting, almost entirely. Tesseract is trained on printed type, and cursive in particular produces nonsense. Handwriting recognition is a different technology.

Tables and multi-column layouts lose their structure. The engine reads in visual order, so a two-column page often comes out with the columns interleaved line by line. Extracting one column at a time by cropping is usually faster than repairing the result.

Unusual typefaces, very small print, low contrast, watermarks over text, and text printed over images all reduce accuracy sharply.

And OCR output always needs proofreading. The characteristic errors are subtle — `rn` read as `m`, `0` as `O`, `5` as `S` — which means a figure in a financial document can be wrong in a way that looks entirely plausible. Never use OCR output for anything consequential without checking it against the original.