Extract Text From an Image

What OCR actually does, and where it struggles

Optical character recognition matches shapes in an image against learned patterns for characters, and the accuracy depends entirely on how clear those shapes are. It performs well on a clean printed document photographed flat, in even light, with a standard font. It struggles on a photograph at an angle, low light, a decorative or unusual typeface, text over a busy background, or a low resolution capture. The single biggest factor is usually the resolution of the text within the image, not the quality of the recogniser, which is why a higher resolution photo of a receipt extracts far more reliably than a close up of the same receipt.

NoteThe biggest factor is pixels per character, not the recogniser. A higher resolution capture extracts more reliably.

Small text is where it fails first

Accuracy degrades as the text gets smaller, and it degrades quietly, which is the dangerous part. A recogniser given text it cannot read returns its best guess rather than an error, and a plausible looking word is indistinguishable from a correct one to the reader skimming the output. Small print on a contract, a serial number, a bank reference, a dosage on a label, and a measurement are all cases where a wrong guess is worse than no answer. For those, verify against the original image character by character, or retype them. Never copy an extracted number into something that matters without checking it.

• Accuracy falls as text gets smaller, and it fails quietly rather than erroring.
• A recogniser returns its best guess, not a warning.
• Small print, serial numbers, references, and dosages are the risk cases.
• Verify extracted numbers character by character before using them.
• Retype critical values rather than trusting the extraction.

NoteExtracted numbers fail silently. Never copy one into anything that matters unchecked.

Getting a much better result from the same image

Most disappointing extractions are caused by the capture rather than the tool. Photographing a page flat, in even light, with the camera parallel to the paper, removes most of the problems at once. Cropping tightly to just the text increases the pixels per character, which is the single biggest improvement available. Rotating an image that is sideways before extraction helps, and increasing the resolution of the original photo helps more than anything. If a document is available as a PDF rather than a photograph, that is a better input than any photograph of it.

• Photograph flat, in even light, with the camera parallel to the page.
• Crop tightly to the text. More pixels per character is the biggest gain.
• Rotate a sideways image before extracting.
• Use a higher resolution original where one exists.
• A PDF is a better input than a photograph of the same document.

NoteCropping tightly to the text gives the biggest accuracy improvement, because it raises the pixels per character.

The privacy advantage, and its limits

Running recognition in the browser means the image is never uploaded, which is a genuine and significant advantage when the material is a payslip, a passport, a medical form, a contract, or anything else you would not want on a third party server. That is the strongest argument for a local tool. The limitation is that local recognition is generally less capable than a large server based service, particularly on difficult images and on uncommon scripts. So there is a real trade: better accuracy against sending a sensitive document to someone else. For sensitive material, the privacy usually wins.

• Local recognition never uploads the image, which matters for sensitive documents.
• Payslips, passports, medical forms, and contracts stay on the device.
• Local recognition is usually less capable on difficult images than a large service.
• For sensitive material, the privacy usually outweighs the accuracy difference.