Copy Text Out of a Photo or Scan (Android)

To get the words out of a photo of a page, run optical character recognition — and the accuracy you get is decided almost entirely at capture time, not by the software. The three things that matter are resolution (aim for text at least 20 pixels tall, which for body text means photographing the page so it fills the frame), contrast (even lighting, no shadow across the page), and skew (text lines within a few degrees of horizontal). Get those right and recognition on clean printed text is routinely above 98% accurate. Get them wrong and no engine rescues you.

On Android, Convertlix's OCR & Text Suite pulls text out of photos and scans on the handset, so the image never leaves the device — no upload, no server-side copy, nothing to retain. The scope, up front: Android 8.0 or later, sign-in required (there is no guest mode), 25 MB per file on the free tier against a server-set monthly allowance of conversions shown in the app, and the output lands in Downloads. Everything below applies whichever engine you use.

How recognition actually works, and why that predicts the failures

OCR is a pipeline, and each stage has a specific way of breaking. Knowing them tells you what to fix.

1. Binarisation. The image is reduced to text-versus-background. Uneven lighting is fatal here: a page with a bright half and a dim half will have one half thresholded into solid white and the other into mush. This is the single most common cause of a bad result.
2. Layout analysis. The engine finds text blocks, columns and reading order. This is where multi-column pages, sidebars, headers and tables cause scrambling — the characters are read correctly and delivered in the wrong order.
3. Line and character segmentation. Lines are found on the assumption that text runs horizontally. A few degrees of skew is tolerated; past roughly 5–10° accuracy drops off, and a keystoned page — one photographed at an angle, so the line spacing changes down the page — confuses it further.
4. Classification. Each character shape is matched against a model. This is where resolution bites: below about 20 pixels of glyph height, the shapes that distinguish rn from m, 1 from l, and 0 from O are simply not present in the image.
5. Language modelling. Results are checked against a dictionary and character probabilities. Helpful on prose, actively harmful on serial numbers, part codes and passwords, where a "correction" turns a right answer into a wrong one.

The practical upshot: straighten and light the page before you recognise it. If the photo is tilted or keystoned, Straighten a Crooked Photo of a Document is the step before this one, and it will measurably improve the result. If you are still holding the paper, Scan a Document to PDF Offline (Android) covers capturing it properly in the first place.

What 300 DPI means when you're using a camera

Scanner guidance says 300 DPI, which is unhelpful advice for a phone. Translate it into framing instead.

An A4 page at 300 DPI is 2,480 × 3,508 pixels. Any phone camera from the last eight years shoots more than that, so long as the page fills the frame. Photograph an A4 sheet from across the room and the page occupies maybe 800 pixels of a 4,000-pixel image — you have thrown away three quarters of your effective resolution before the engine sees it.

So: fill the frame with the page, and check the glyph height rather than the megapixels. The rule of thumb worth remembering is that body text needs about 20 pixels of height to be reliably recognisable. If you can zoom into your photo and read the smallest text comfortably, so can the engine. If it looks soft to you, it will be guessing.

Two capture details that pay off more than any setting: photograph in even, indirect light rather than under one bright lamp, and hold the phone parallel to the page. Both are geometry, not equipment.

OCR does not make a PDF smaller

This trips people up regularly, so it's worth stating flatly: recognition adds to a scanned PDF. It produces a text layer positioned invisibly over the existing page images, which is what makes the document searchable and selectable — the page images are still there at full size underneath. The file gets very slightly larger.

Recognition and compression are separate jobs with separate tools. If the document is oversized, that is an image-encoding problem: Why Is My Scanned PDF So Large? explains where the bytes are, and Compress a PDF Under 2 MB on Android (Free) is the lever. Do the compression after recognition if you need both, and don't compress so hard that you destroy the character shapes the text layer was built from.

How to get accurate text out of a photo

1. Fix the capture before you fix anything else. Even indirect light, no shadow across the page, phone parallel to the paper, page filling the frame. If you can retake the photo, retaking it beats every other step on this list.
2. Straighten and crop to the text area. Level the page using a line of text as your reference, and crop away the table it is sitting on. Less non-text area means fewer things for layout analysis to misinterpret.
3. Handle columns and tables separately. If the page has two columns, crop and recognise each one on its own. This is the difference between usable text and interleaved nonsense, and no setting fixes it.
4. Run the recognition. In Convertlix, open OCR & Text Suite and pick the image through Android's picker; the text is extracted on the phone.
5. Proofread the things a dictionary can't help with. Numbers, codes, names, dates, currency amounts and units. The language model that fixes ordinary words is exactly what corrupts a serial number, so check those character by character.
6. Keep the image. The extracted text is a derivative. If a figure looks wrong later, the photo is your source of truth.

When this doesn't work

• Handwriting. Printed-text OCR and handwriting recognition are different problems. Very neat block capitals sometimes come through; ordinary cursive does not, and treating a result as reliable when it isn't is worse than not trying.
• Low-resolution sources. A screenshot of small text, a photo taken from too far away, or an image someone already downscaled and sent you over chat. The detail isn't there to recover.
• Complex layouts. Forms, invoices with nested tables, newspapers, and anything where reading order carries meaning. You will get the characters and lose the structure.
• Stylised or decorative type. Logos, script fonts, condensed display faces and text over a photograph all defeat the classifier.
• Faded, creased or stamped documents. A stamp over text, a fold through a line, or carbon-copy paper removes the contrast the whole pipeline depends on.
• Anything where an error is expensive. Do not trust extracted text for prescriptions, legal figures, bank details or exam answers without reading it against the original. OCR is a time-saver, not an authority.

FAQ

How accurate is phone OCR?

On clean printed text photographed well — even lighting, page filling the frame, lines close to horizontal — recognition of body text is routinely above 98% accurate, which means a handful of characters to fix on a page. Accuracy falls sharply with shadow across the page, skew beyond about 5–10°, and text below roughly 20 pixels of glyph height. Handwriting, decorative type and multi-column layouts are separate problems where the number is much lower and much less predictable.

Does OCR make a scanned PDF smaller?

No — it makes it very slightly larger. Recognition adds an invisible text layer over the existing page images so the document becomes searchable and selectable; the images themselves are untouched and still account for nearly all the file size. Reducing the size of a scanned PDF is a separate operation that re-encodes those images, and it is worth doing after recognition rather than before.

Why is the extracted text in the wrong order?

Because the page has a layout the engine's analysis stage read differently from you — two columns, a sidebar, a pull quote, or a table. The characters are usually recognised correctly and then assembled in the wrong reading order. Cropping the image to one column at a time and recognising each separately fixes it reliably, and is faster than untangling the output.

Does Convertlix upload my photo?

No. The image is opened through Android's own file picker, the text is recognised on the phone, and nothing is sent anywhere — and the app has no storage permission, so it cannot read anything you did not hand it. Its network calls carry identity, licence, quota, configuration and payment data, never an image. Sign in once, switch on airplane mode, and recognition still works, which is unusual in this category: many OCR features send the image to a server to run the model.

Is there a free way to extract text from an image?

Yes, several, and some are excellent. Google Lens is free, built into Android and very good at this — it sends the image to Google's servers to do it, which is the trade. Microsoft Lens and Adobe Scan both include free OCR with cloud processing. Offline and free on a computer, Tesseract is the open-source reference engine (tesseract page.jpg out -l eng) and handles whole folders. As of September 2026 the practical difference between these options is rarely accuracy on clean printed text — it is whether the picture of your document goes to somebody's server, which for a payslip or a medical letter is worth deciding deliberately.

---

If the page is a photo on your phone and you need the words rather than the picture, straighten it, fill the frame, and read the numbers back against the original.