Most of the result is decided before the shutter. Here is how to shoot it, what a crooked or badly lit frame costs, and how to get it into a sheet without uploading it.
Quick answer:
Convert the photo with image to PDF, then extract it with Read scanned pages (OCR) ticked in the table extractor. Both steps run on your device. Shoot parallel to the paper, directly above, in indirect daylight, with the page flat and the table filling the frame — and if your phone has a document scanner, use it first, because it corrects the perspective and writes a PDF directly. Measured cost of getting it wrong: rotating a five-column table 1.5° took cells exactly right from 100% to 91%, though all five columns survived either way.
How should I take the photo?
Parallel to the paper, directly above the middle, in indirect daylight, page flat, table filling the frame. Avoid shooting under a ceiling light — your phone throws its shadow across the page.
What does a crooked photo cost?
Measured on a five-column table: 0.5° left cells exactly right at 100%, and 1.5° brought it to 91%. All five columns survived either way — structure is tough, characters are not.
Should I use my phone's document scanner?
Yes, if it has one. Notes on iPhone and Google Drive on Android correct the perspective and flatten the lighting, and write a PDF directly — better input than a raw photograph, and one step fewer.
Can it read handwriting?
No. The engine reads printed characters, so a handwritten ledger is effectively unreadable rather than merely inaccurate. Worth knowing before photographing forty pages of one.
Is the photo uploaded?
No. Converting it and recognising it both happen in your browser — which matters, since a photographed document is often the most sensitive thing on a phone.
Both steps run on your device. No app, no account, no upload.
Photograph it properly and the rest is easy
Almost everything that determines the result is decided in the two seconds before the shutter, not by any software afterwards. A camera introduces four problems a scanner does not have, and all four are yours to avoid.
Angle
Hold the phone parallel to the paper, directly above the middle. Leaning over the page at an angle makes the far edge narrower than the near one, so the columns are no longer parallel lines — they converge. Column detection works on vertical gaps, and gaps that lean do not line up down the page.
Rotation costs measurably too. On a five-column table rendered at 150 dpi, rotating by 0.5° left cells exactly right at 100%; at 1.5° that fell to 91%. The columns themselves survived both — 5 of 5 — which is the pattern throughout: structure is tough, characters are not.
Light
Daylight, indirectly, with the page out of your own shadow. The worst common case is a phone held directly overhead under a ceiling light, which throws the phone's shadow across the middle of the page — right where the columns are. Glossy paper adds a specular highlight that whites out whatever is under it. Move, rather than turning the flash on: flash produces a bright centre and dark corners, which is worse than the shadow it replaced.
Flatness
A page curling out of a binder, a receipt still curved from the till, a folded statement — the curve bends the lines of text and darkens the fold. Flatten it under something, or at minimum press the fold open. This is the one problem no setting fixes.
Distance and focus
Fill the frame with the table and nothing else. A photograph of a desk with a document on it spends most of its pixels on the desk, and the characters get whatever is left. Tap to focus before shooting, and check the result is sharp before you walk away — a blurred photograph cannot be un-blurred.
Getting it in: two steps
Image to PDF — your phone's JPG or HEIC becomes a PDF. HEIC is handled directly; an iPhone photo often arrives with no file type set at all, which is fine.
Both run in your browser. A photograph of a document is frequently the most sensitive thing on a phone — a payslip, a medical form, a client's statement — and it never leaves the device here.
Before either step, use your phone's own document scanner if it has one. The Notes app on iPhone and Google Drive on Android both detect the page edges, correct the perspective and flatten the lighting, and their output is significantly better input than a raw photograph. They write a PDF directly, so you can skip step one entirely.
What a photograph costs against a scan
Our quality sweep gives the shape of it. All five columns of a five-column table survived every condition tested, including 56 dpi with heavy grain and 1.5° of rotation. Cells exactly right is where the input quality shows: 100% at 150 dpi clean, 91% at 100 dpi, 85% at 56 dpi, 97% with heavy grain and hard compression.
A good phone photograph of an A4 page sits in the upper part of that range; a hurried one at an angle in poor light sits at the bottom. So the realistic expectation is: you will get your columns, and you should check your figures. On a dense 30-row page even a good capture is a page worth reading over — at 150 dpi equivalent, 60 of 60 money figures were exact, and on a rough version of the same page 53 of 60.
Cells that cannot be what their column is are marked ⚠ for you — a capital O inside a money figure, for instance. That is a narrow check by design: 6 of 6 on deliberately corrupted figures with no false alarms, nothing raised on 6 correctly extracted typed pages, and roughly one OCR error in five caught overall. The rest are in text columns where there is nothing to check a word against.
Photographing several pages
Shoot them all in one session, in the same place, with the same framing — consistency matters more than perfection, because the same lighting and the same distance give the same character size throughout. Then put them in order in image to PDF, one per page, and extract once.
If the pages are one continuous table, choose Merge matching tables and they become a single sheet with the headings kept once. A page is treated as continuing the previous one when the column count matches, the headings match or are absent, and the columns still sit where they did — over ten multi-page fixtures that got 10 of 10 right with nothing wrongly merged. A photographed set is where you should glance at the preview before trusting the join, since small differences in framing move the columns slightly between shots.
The cases where a photograph will not work
Handwriting. The engine here reads printed characters. A handwritten ledger is effectively unreadable, not merely inaccurate — worth knowing before you photograph forty pages of one.
A screen photographed with a camera. You get moiré banding and the pixel grid fighting the camera's sensor. Take a screenshot instead; see image to Excel.
A page at a sharp angle. Perspective is not corrected here. Reshoot it square, or run it through your phone's document scanner first, which does correct it.
Very small print in a low-light photo. Characters a few pixels tall carry no information to recover. Get closer, get more light, or find the original document.
A language other than English. Recognition here is English only, and an unfamiliar language is read badly rather than refused.
Where to go next
Image to Excel — for a screenshot instead, which is the best input this route ever gets.
Receipt to Excel — the commonest thing anyone photographs, with the columns an expense claim needs.
Scanned PDF to Excel — if you have access to a scanner after all — it is worth the walk.
Frequently Asked Questions
Convert the photo to a PDF with image to PDF, then drop it into the PDF table extractor with Read scanned pages (OCR) ticked. Both steps run in your browser, so the photograph never leaves your phone or laptop. If your phone has a document scanner — Notes on iPhone, Google Drive on Android — use it first: it corrects the perspective and flattens the lighting, and writes a PDF directly.
Parallel to the paper and directly above the middle, in indirect daylight, with the page flat and the table filling the frame. Avoid shooting under a ceiling light directly overhead, because your phone throws its own shadow across the middle of the page. Do not use flash — it produces a bright centre and dark corners, which is worse than the shadow it replaced.
Measured on a five-column table: rotating it 0.5 degree left cells exactly right at 100%, and 1.5 degrees brought that down to 91%. The columns themselves survived both — 5 of 5 — which is the pattern throughout: the structure is tough and the characters are not.
Almost always. Across every condition we tested on a five-column table — down to 56 dpi with heavy grain, and at 1.5 degrees of rotation — all five columns came back. Cells exactly right is where quality shows: 100% at 150 dpi clean and 85% at 56 dpi. Expect to get your columns and to check your figures.
No. The engine reads printed characters, and a handwritten ledger is effectively unreadable rather than merely inaccurate. That is worth knowing before photographing forty pages of one.
You can, but do not. Photographing a screen gives moiré banding where the pixel grid fights the camera's sensor, plus reflections and uneven brightness. A screenshot is pixel-exact and is the best input this whole route ever gets — see image to Excel for that path.
Shoot them in one session with the same framing and lighting, since consistency matters more than perfection — the same distance gives the same character size throughout. Put them in order in image to PDF, one per page, and extract once. If they are one continuous table choose Merge matching tables; over ten multi-page fixtures that got 10 of 10 right with nothing wrongly merged. Glance at the preview, because small differences in framing move the columns between shots.
No. Converting it to a PDF and recognising it both happen in your browser. A photograph of a document is often the most sensitive thing on a phone — a payslip, a medical form, a client's statement — and it stays on the device. The only network request is for the recognition engine, about 6 MB the first time, which is then cached.
That is the intended route: photograph the table, convert it with image to PDF, extract it, all in the phone's browser with nothing installed and nothing uploaded. Two honest caveats — a phone has far less memory than a laptop, so a very large image or a long document can fail where a laptop copes, and correcting a cell means double-tapping it, which is fiddly on a small screen.
It is a web page, so it runs in the browser on both, and iPhone HEIC photos are handled directly — they often arrive with no file type set at all, which is fine. Use your phone's own document scanner first if it has one: Notes on iPhone and Google Drive on Android both correct the perspective and flatten the lighting, and write a PDF directly, which skips a step and gives better input than a raw photograph.
Only if the writing is printed rather than handwritten, which on a whiteboard it rarely is — the engine here reads printed characters and handwriting is effectively unreadable rather than merely inaccurate. A projected or printed table photographed off a wall can work, though glare from the surface is the usual problem.
You can, and the binding is the difficulty. A page curving into the spine bends the lines of text and darkens the gutter, and perspective is not corrected here. Press the page as flat as it will go, shoot square from directly above, and if the inner column still comes out wrong, photograph the halves separately and extract each.
As many as you like — put them in order, one per page, in image to PDF and extract once. Shoot them in a single session with the same framing and lighting, because consistency matters more than perfection: the same distance gives the same character size throughout, and small differences in framing move the columns between shots, which is what the grouping compares.