Text trapped in a picture cannot be searched, corrected, translated or copied. Upload a PNG or JPG above and OCR reads the characters out of the image, then builds the result as a Word document, PDF, PowerPoint file or plain text.
OCR is genuinely useful and genuinely imperfect. The single biggest factor in your result is not the software — it is the image you feed it. Below: how to capture a page that converts almost perfectly, the mistakes that are most dangerous, and how to check the output properly.
Five minutes of care here saves an hour of correcting afterwards. Recognition depends on the software being able to see clean, separated letter shapes.
If you are scanning rather than photographing, 300 DPI is the standard target for text. Below about 200 accuracy drops noticeably; above 600 you gain nothing but file size.
JPEG compression was designed for photographs. It discards detail the eye is unlikely to miss in soft, gradual tones, and it handles hard edges badly — while text is nothing but hard edges.
The result is faint halos and ringing around letters. Your eye compensates without noticing. OCR does not: those artefacts sit exactly where the software is trying to find the boundary of a character, and they are a real source of misreads in small type.
For screenshots, always capture as PNG. It is lossless, every pixel survives, and recognition is measurably better. For photographs of paper, JPEG is unavoidable from most cameras — just use the highest quality setting and do not re-save the file repeatedly, since each save compounds the damage.
One thing that cannot be undone: converting an already-compressed JPEG to PNG does not restore lost detail. It preserves the damage in a bigger file.
OCR expects text to run horizontally. Even a few degrees of rotation reduces accuracy, and a sideways page can fail almost completely.
Most engines attempt to detect and correct orientation automatically, but detection is itself a guess and it works less well on pages with sparse text, heavy tables, or non-Latin scripts. Rotating the image yourself removes the uncertainty entirely.
The same applies to skew — a page that entered a scanner crooked. A three-degree tilt is enough to hurt line detection, which is what determines where one line of text ends and the next begins.
Recognition is pattern matching, so the errors follow predictable shapes. Certain characters are genuinely ambiguous at small sizes even to a human eye:
0 and O · 1 and l and I · 5 and S · 8 and B · 2 and Z · rn read as m · cl read as d
In prose these are usually obvious and quick to fix — a misspelled word stands out. In anything numeric they are dangerous, because a wrong digit still looks like a plausible number. An account code, an invoice total, a dosage, a serial number or a date can be silently wrong and nothing will flag it.
So proofread proportionally to the stakes. A quote you are pasting into an essay needs a quick read-through. A table of figures you intend to calculate with needs checking against the original, value by value. Never treat OCR output of financial, medical or legal data as verified.
OCR recovers text, not design. Columns, precise spacing, decorative typography and exact positioning are not reconstructed, because the image contains no information about them beyond how things happen to look.
Two consequences worth expecting:
Multi-column pages can interleave. A newspaper or journal page with two columns may come out with lines alternating between them, since the software reads across the page. Cropping each column separately and converting them one at a time gives a far better result than fighting the output afterwards.
Tables lose their grid. Values are recovered but the row and column relationships usually are not. For a table you actually need to calculate with, extract the text first and rebuild the structure in a spreadsheet, or start from a PDF rather than an image where possible.
Choosing the output format is really choosing what you want to do next: Word for editing and correcting, plain text for pasting into something else, PDF for a searchable archive copy, PowerPoint for building slides from a captured page.
Standard OCR is trained on printed type. Handwriting is a different problem entirely — recognising it reliably requires different technology, and even then cursive and individual quirks defeat it. Expect poor results from handwritten notes, signatures and annotations.
Also difficult: heavily stylised display fonts, text over busy photographic backgrounds, very low contrast such as grey on grey, faded thermal receipts, watermarked pages where the mark crosses the text, and text smaller than roughly 8 points in the original.
Working with a scanned PDF rather than an image? OCR PDF handles multi-page documents. Rotate PDF straightens sideways scans before recognition, and Image to PPTX places a picture on a slide without extracting its text. Browse all image tools.