Why OCR Gets Text Wrong and How to Improve Accuracy on Screenshots, Scans and Photos

OCR first finds the layout: blocks of text, lines and characters. Then it recognizes each character shape and uses a language model to choose between similar candidates. Two things hurt both stages: small, blurry or noisy characters that make shapes ambiguous, and complicated layouts such as columns, tables and text over pictures that confuse the line detection.

• ▸Resolution: aim for capital letters at least 20 to 30 pixels tall. For a scanned page, 300 dots per inch is the usual target. Going far beyond that adds file size without benefit.
• ▸Screenshots: take them at the screen's native size or zoom the page in before capturing, rather than enlarging a small screenshot afterward. Enlarging blurs the edges.
• ▸Photos of paper: hold the camera parallel to the page, use even light without shadows or glare and keep the whole page in frame. Tilt and curved pages distort lines.
• ▸Avoid flash on glossy paper. Use daylight or a diffuse lamp.
• ▸Save as PNG for screenshots and scans of text. Heavy JPEG compression creates artifacts around letters that look like extra marks.

• ▸Crop to the text you need. Borders, page edges and photos next to the text produce junk lines.
• ▸Straighten the page. Even a tilt of a few degrees lowers accuracy.
• ▸Convert to grayscale and raise the contrast so that dark text is clearly separate from a light background.
• ▸Remove background colors and textures, and dark scanner borders.
• ▸Invert light text on a dark background so that the text is dark on light, which most engines expect.
• ▸Split multi column pages into separate images, one per column, so the lines are not read straight across.

• ▸Select the language of the text. The engine uses language data to disambiguate characters, and the wrong one produces wrong words.
• ▸Mixed language text is harder. Process each part separately if you can.
• ▸Plain printed text in common fonts works best. Decorative fonts, very small print and stylized logos are unreliable.
• ▸Handwriting is not supported by general purpose engines, apart from very neat block letters.
• ▸Tables come back as lines of text. Recognize columns individually or rebuild the table by hand.

OCR errors are patterned, so your proofreading can be targeted.

• ▸The digit 0 and the letter O, the digit 1 and the letters l and I, and 5 and S.
• ▸The pair rn read as m, and cl read as d.
• ▸Missing or extra spaces, and hyphenated words split at line ends.
• ▸Quotation marks and apostrophes read as other symbols.
• ▸Dropped diacritics on accented letters, especially at small sizes.
• ▸Numbers, which are the most expensive errors. Verify every figure, date and code against the original.

Capture at a good size, crop and straighten, select the language and run the Image to Text Converter. It works in your browser and also accepts a pasted screenshot. Read the result next to the original, fix the pattern errors above and run Remove Line Breaks to rejoin paragraphs if the source was a printed page. For anything legal, financial or medical, proofread every number and name by eye. Recognition saves typing, but it does not replace checking.