How to Extract Text from an Image, OCR Explained

OCR stands for Optical Character Recognition. It is the technology that analyses an image and identifies the shapes of characters, letters, numbers, punctuation, converting them into actual text data that can be copied, searched, and edited. OCR powers everything from bank check processing to Google's indexing of scanned book pages.

Modern OCR engines are trained on millions of text samples across many fonts, styles, and layouts. Accuracy on clean, printed text in common fonts from standard document sources exceeds 99% in most cases. The remaining errors are concentrated in edge cases, unusual fonts, damaged text, extreme skew, which is why verifying the output of critical documents remains the correct practice even with a high-accuracy engine.

For documents with mixed content, such as a scanned report that contains both text paragraphs and data tables, treat the OCR output as a starting point rather than a final product. Tables in particular require post-processing to restore the row and column relationships that the OCR engine outputs as a flat sequence of text tokens.

OCR engines trained primarily on Latin script may misread characters from other writing systems, even when those characters appear alongside Latin text in the same image. If your document mixes scripts, bilingual menus or invoices with mixed alphabets for example, check the extracted text for each script separately rather than reviewing the combined output as a single block.

You scanned a physical document and need to edit the text. You received a PDF that is actually a scanned image and the text is not selectable. You have a photo of a sign, whiteboard, receipt, or business card and need the text in digital form. You want to digitize handwritten notes, though accuracy on handwriting is significantly lower than on printed text.

For handwritten notes in particular, accuracy drops dramatically below the level useful for direct editing. Handwriting OCR is most practical for capturing approximate content for search or reference, not for producing an exact transcript. If accuracy on handwriting matters, a specialized handwriting recognition service trained specifically for that use case produces better results than a general-purpose OCR engine.

Batch OCR workflows benefit from preprocessing images consistently before recognition. Applying the same contrast, brightness, and deskew corrections to every image in a batch, rather than optimising each image individually, produces more uniform output. Consistent preprocessing reduces the variance in recognition quality across the batch and makes post-processing more predictable.

Upload your image to the Image to Text tool. The tool runs Tesseract.js, an open-source OCR engine, entirely inside your browser. No image data is transmitted anywhere. After processing, the extracted text appears in a scrollable text area. Use the Copy button to copy all extracted text to your clipboard in one click, then paste into any document or editor.

The output text from OCR is plain text without formatting. Bold, italic, font sizes, and table structures are lost during the recognition step. If you need to preserve some of the document's visual structure, headers, bullet points, table cells, consider a PDF-to-Word conversion tool that combines OCR with layout analysis to approximate the original document's formatting in an editable word-processing format.

OCR accuracy depends heavily on input quality. Sharper, higher-contrast images produce better results. Before running OCR, try: ensuring the image is well-lit and in focus, straightening any skew or rotation, converting to grayscale to improve contrast, and cropping away borders and surrounding clutter. Printed text in a standard font at 300 DPI or higher is ideal.

Skew correction, rotating the image to make text lines horizontal, is one of the most impactful preprocessing steps for OCR accuracy. Even a 3-5 degree tilt can significantly reduce accuracy by breaking the assumptions the recognition model makes about text line geometry. Most document scanner apps include automatic deskew; if running OCR on a manually photographed document, deskew the image before uploading.

For ongoing digitisation projects, establish a quality control step where a sample of extracted text is compared to the source image after each batch. Spot-checking ten percent of pages catches systematic errors, such as a consistently misrecognised character or a font that the engine struggles with, before those errors propagate through the entire dataset.

Handwriting accuracy is unpredictable, highly stylized or cursive writing may produce garbled results. Unusual fonts, decorative text, and very small text challenge OCR engines. Complex layouts with multiple columns, tables, and wrapped text may produce out-of-order extracted text that needs manual cleanup. Languages with non-Latin scripts require a model trained on that script, verify language support before relying on results.

For documents in non-Latin scripts, Chinese, Arabic, Hindi, Japanese, Korean, OCR requires a language-specific model trained on those characters. Applying an English-language OCR model to Arabic text will produce meaningless output. Tesseract, the engine used by this tool, supports over 100 languages, but you need to verify that the relevant language pack is enabled for the language of your document.