How OCR Works and How to Get Better Text From Images

You photograph a page and want the words as editable text, or you have a scanned document that cannot be searched. Optical character recognition, or OCR, is the technology that reads letters from pictures. It can be impressively accurate or frustratingly wrong, and the difference usually comes down to the picture, not the software. This guide explains what happens inside OCR and what you can do to get better results.

What OCR does

A picture of text is just a grid of coloured pixels. OCR analyses that grid, works out which shapes are letters, and outputs the characters they represent as plain text. Modern engines such as Tesseract, the open-source engine used by the OCR - Image to Text tool, use neural networks trained on huge numbers of examples of text in many fonts and languages.

The steps inside OCR

1. Preparation. The image is converted to black and white or greyscale, contrast is adjusted, noise is removed and the page is straightened if it is tilted.
2. Layout analysis. The engine finds blocks of text, then lines, then words, and works out the reading order, including columns.
3. Recognition. A neural network reads each line as a sequence, predicting the most likely characters. Language data helps it choose between similar shapes, such as an l, a 1 and an I.
4. Post-processing. Dictionaries and language rules correct likely errors and the result is assembled into text.

Because the language matters, choosing the correct language before you run OCR improves accuracy.

Why results vary so much

• Resolution. Letters need enough pixels to be distinct. Text less than about 20 pixels tall is unreliable, and 30 or more is better. For a printed page, a 300 dpi scan is the usual target.
• Contrast and lighting. Dark text on a light, plain background works best. Shadows, glare and coloured backgrounds cause errors.
• Angle and curvature. A page photographed at an angle or a curved book page distorts letters.
• Font and content. Common printed fonts work well. Decorative fonts, italics at small sizes and handwriting are much harder.
• Layout. Tables, columns and text over pictures confuse reading order.

A checklist for better results

1. Shoot flat and straight. Lay the page on a table and photograph it from directly above, with the lines parallel to the frame.
2. Use even light. Daylight from a window is ideal. Avoid the flash, which causes glare on glossy paper.
3. Fill the frame. Move closer so the text is large, but keep the whole page inside the image.
4. Focus. Tap to focus and hold steady. Blur ruins recognition.
5. Crop. Remove everything except the text using the Image Cropper, so borders, backgrounds and pictures are not read as text.
6. Adjust size. If the image is small, enlarge it with the Image Resizer; if it is huge, reduce it to a sensible size.
7. Pick the language before you run the recognition.
8. Proofread the result. Never trust OCR output for names, numbers or dates without checking.

Typical errors to look for

OCR mistakes follow patterns. Watch for 0 and O, 1 and l and I, 5 and S, rn read as m, cl read as d, and missing or extra spaces. Hyphens at line ends may split words, and page numbers or headers may appear in the middle of the text. Numbers in tables and prices deserve special care, since a dictionary cannot help with them.

Scanned PDFs

A scanned PDF is usually a set of page images with no text inside. To get the text, convert the pages to images with PDF to JPG at 200 to 300 dpi, then run each image through OCR. If your PDF already contains selectable text, you do not need OCR; convert it with PDF to Word instead. To turn photographs of pages into a single document, use JPG to PDF.

Handwriting, tables and other limits

Engines like Tesseract are designed for printed text. Handwriting recognition needs specialised models and is far less reliable, so expect to correct most of the output. Tables and columns are read as lines of text, so the structure is lost and you may need to rebuild it by hand. Very long or very high-resolution images can be slow on older devices.

Making the most of the text

Once you have the text, check it with the Grammar Checker to catch mistakes OCR left behind, condense long passages with the Text Summarizer, and count words with the Word Counter. For advice on how long different kinds of text should be, read the word count guide.