A photo of a page is useless to software until something turns the pixels back into text. The Docyield image extractor does exactly that. Upload a PNG, JPG, or WEBP — a phone photo of a document, a screenshot, a scanned page — and it reads the text and data inside and hands it back as clean text, or as structured JSON, CSV, or XML when you need fields and tables rather than prose.
An image carries no text at all, only a grid of coloured dots that happen to form letters to a human eye. There's nothing to select, copy, or search until those shapes are recognised as characters. The extractor runs optical character recognition to recover the text, then interprets the layout — reading order, columns, tables — so the output reflects what the image actually says instead of a scrambled stream of words.
Turning pixels into usable text
When you upload an image, Docyield runs OCR to identify the characters, then reconstructs how they're arranged on the page. It keeps paragraphs together, follows multi-column layouts in the right order, and preserves tables as rows and cells rather than collapsing them into a run of stray numbers. The plain-text result is the image's content, ready to search, quote, edit, or paste somewhere it's actually usable.
Beyond plain text, the extractor can return structured data. Where an image contains labelled fields or a table, those become keys and arrays in JSON, rows in CSV, or elements in XML — so you can consume the content programmatically instead of scraping a text blob afterwards.
Why OCR-and-structure beats retyping
Faced with an image of text, most people just retype it — slow, dull, and a fresh chance to mistype a figure with every line. Basic OCR removes the typing but stops at characters: it gives you the words without telling you which is a label, which is a value, or where a table's rows begin.
The image extractor goes a step further. It preserves reading order, keeps tables intact, and — when you ask for structured output — maps fields into keys and repeated rows into arrays. You get content you can use directly rather than text you still have to untangle, and you skip the error-prone manual transcription entirely.
Screenshots, signs, and snapshots
The images people need read fall into a few recurring buckets, and each behaves a little differently. Screenshots are the easiest: they are crisp, perfectly aligned, and lit by the screen itself, so the text comes through almost verbatim — handy for lifting an error message, a chat log, or a table from an app that offers no export. Photographs of printed pages are the next step up in difficulty, since focus, angle, and lighting now come into play.
Then there are the harder, real-world captures: a sign across a room, a label at an angle, a receipt curling on a desk. These can still be read, but they sit at the limit of what OCR manages reliably, and a value or two may need confirming. Knowing roughly which bucket a given image falls into sets a fair expectation for the result and tells you whether a quick review is worth doing before you rely on the output.
Who extracts text from images
- People capturing documents on a phone who need the text rather than a picture.
- Support and ops teams pulling details out of screenshots customers send in.
- Developers feeding image content into search, LLMs, or data pipelines.
- Researchers and students digitising photographed pages, slides, and notices.
- Anyone who has text locked in an image and wants it editable and searchable.
Accuracy and the limits of an image
OCR quality is driven by image quality. A sharp, well-lit, straight-on photo or a clean scan reads reliably; a blurry, dim, or steeply angled shot is genuinely harder, and very low resolution or heavy glare can defeat even good OCR. Filling the frame with the page, holding the camera level, and avoiding shadows all make a clear difference to the result.
The extractor doesn't fabricate content. Where a region is illegible or a value simply isn't there, it returns nothing for that part rather than a confident guess, because a blank is easier to catch and correct than a wrong answer wrapped in plausible text. For the small share of images that are awkward or low quality, a quick check against the original is the cheapest safeguard. Handwriting in particular is far harder than printed text and benefits from review.
Layouts, languages, and tables in images
Photographed and scanned pages come with the same complications as any document — multiple columns, headers and footers, sidebars, and tables sitting beside prose. Read in raw position order, these produce text that hops between columns mid-sentence. Docyield reconstructs the reading order so multi-column images come out coherent, and keeps tabular regions as rows and cells.
The extractor reads a wide range of languages and scripts, not just English, so images of non-Latin or mixed-language text are handled rather than mangled. Where an image holds both prose and a table, both are preserved instead of one being lost to the other.
Capturing a better photo
Because OCR is only ever as good as the picture it's given, a little care at capture time pays back more than anything you can do afterwards. Lay the page flat and shoot it square-on rather than at an angle, so the lines of text stay parallel instead of converging. Fill the frame with the page so the characters are large and the camera has plenty of detail to work with, and let autofocus settle before you take the shot — motion blur is one of the most common reasons text comes back garbled.
Lighting is the other half. Even, diffuse light avoids the hard shadows and bright hotspots that swallow whole lines, and turning the page slightly to dodge glare from a lamp or window often rescues a shot that would otherwise be unreadable. If a first capture comes back patchy, retaking it with these adjustments is almost always faster than correcting the output by hand, and it usually fixes the problem at the source.
Output formats and going to volume
Plain text is the default and covers a lot of uses — search, quoting, feeding content to an LLM. The same extraction is also available as JSON, CSV, Excel (XLSX), or XML from the result view, so structured fields and rows are one tab away when prose isn't enough. They all come from a single read of the image, so a second format never means uploading again.
The free extractor handles one image at a time, which suits one-off captures. When you need to process images continuously, in bulk, or straight from your own software, the Docyield API and batch dashboard run the same extraction with webhook delivery, so the behaviour you test here stays identical at scale. The dashboard shows each image's status and lets you re-run any that need a second pass, which keeps a steady stream of captures manageable.
How to extract text from an image
- 1Upload your image — drop a PNG, JPG, or WEBP onto the box above, or click to choose one.
- 2Wait a few seconds while Docyield runs OCR and reconstructs the layout.
- 3Review the extracted content and check any part against the original image.
- 4Choose your output — plain text, or JSON, CSV, Excel, or XML for structured data.
- 5Copy the result or download the file, ready to search, edit, or feed into your tools.
Frequently asked questions
Processing documents at scale?
Batch upload, an extraction API, and webhooks for 100+ documents a month.