The table exists on paper. It is in a ledger, on a printed statement, on the whiteboard photograph from a meeting, on a supplier's price list that was never a file. What you need is the same table inside Excel — where the amounts can be totalled, sorted and checked — and the usual way to get it there is to type it, cell by cell, misreading one figure in fifty and finding out at the totals row.

There is a shorter way. The JPG to Excel converter on this site reads a photograph of a table and rebuilds it as a real spreadsheet: rows and columns restored to a grid, amounts typed as numbers rather than text, and an editable grid to check before the file is created. This guide walks through the whole job — how to photograph the table, how the rebuild works, what to check, and the one limitation of the built-in alternative that matters more than any other.

How do you photograph a table so it converts cleanly?

The conversion can only work with what the photograph gives it, so the first steps are photographic. Lay the page flat and shoot from directly above, the phone parallel to the paper, so the table's rows run level across the frame — a table photographed at a slant has columns that lean, and the rebuild has to guess where each cell begins. Fill the frame with the table and little else. Use even light without glare: a printed page under a bulb photographs with a bright pool in the middle and grey corners, and faint figures in the grey are the ones that get misread. If the table runs over two pages, convert them separately and stack the results in Excel afterwards — one clean page converts better than one photograph of two pages at an angle each.

Screenshots and scans work too, and are usually easier than photographs: a screenshot of a table is already flat, level and evenly lit. However the table reaches you — camera, scanner, screenshot, a picture someone sent you — the same converter takes it from there.

How does the conversion rebuild the table, step by step?

Drop the image in and pick the language the table is written in — more on that choice shortly, because it decides more than people expect. The engine first reads the words together with their positions on the page and a confidence score for each reading. Then it rebuilds the structure: words at the same height are grouped into rows, and the columns are found from the gaps — the vertical channels of empty space that repeat from row to row where one column ends and the next begins. A single line of text can bridge a gap by accident, so the boundaries are set by the gaps that most rows agree on, not by any one row. If a band of the table produced no readable line at all — a dark header row is the usual case — that band is cropped out and read again on its own, so a styled header does not silently vanish.

What comes back is a grid on screen, before any file exists. That order — grid first, file second — is the part of the workflow worth protecting, because the grid is where the checking happens.

Why do the amounts arrive as real numbers, not text?

Because a spreadsheet full of text that looks like numbers is a trap: it displays fine and totals to nothing. So the converter types each cell by what it contains. Figures with currency symbols, thousands commas or decimals become real numeric cells with the value intact — 12.50 in a cell is the number 12.5, ready for a SUM. Amounts in parentheses become negatives, the way accounts print them. Percentages and dates stay as text, deliberately, because converting those guesses wrong would corrupt data instead of merely labelling it. The test of a converted table is the totals row: if the column sums to the figure printed at the bottom of the paper, the numbers are numbers.

What should you check in the grid before you download?

Readings the engine is unsure about are flagged in the grid rather than hidden, so check the flagged cells first — they are where a 5 is really an S, or a worn digit could be two different numbers. Then scan the structure: every printed row present, no two columns merged into one, the header row reading as headers. Anything wrong can be edited in the grid itself before the download; the spreadsheet is generated from the grid you approved, not from the raw reading. When it looks right, download the .xlsx — with the header styled and column widths set — or take a CSV if the next tool in your chain prefers it.

Honesty about the weak case: this reads print. Typed and printed tables, screenshots and clean scans convert well. Handwriting is a different problem — a handwritten ledger converts unevenly, and figures in it deserve cell-by-cell checking rather than a totals glance.

Which languages can it read — and why does that decide the tool?

The converter reads 28 languages, with English as the default and Urdu second in the list, its data hosted on this site so it loads as fast as English. Arabic, Persian, Hindi and Bengali are there too, along with the major European and East Asian scripts. You pick the language before converting, and changing it re-reads the same image — no re-upload.

This is where the obvious alternative runs out. Excel has its own built-in version of this feature, Data from Picture — and Microsoft publishes exactly which scripts it can read:

Only the following character sets are supported: English, Bosnian, Croatian, Czech, Danish, Dutch, Finnish, French, German, Hungarian, Italian, Norwegian, Polish, Portuguese, Romanian, Serbian, Slovak, Slovenian, Spanish, Swedish, and Turkish.
— Microsoft Support, Insert data from picture

Count them: 22 character sets, and Urdu, Arabic, Persian, Hindi and Bengali are not among them. If the table in front of you is written in a script Excel's own tool refuses, the converter here is not the alternative to the built-in feature — it is the only version of the job that runs at all. The image to Excel converter is the same engine under its general name, for pictures that did not start life as JPGs.

What about invoices and bank statements?

An invoice is a table with a letterhead attached, and the converter treats it that way: the lines above the table become rows of their own, the itemised part rebuilds as the grid it is, and the amounts arrive as numbers you can reconcile against the printed total. A bank statement is the special case with its own tool — the bank statement converter — because statements arrive as PDFs, often multi-page, and because their columns (date, description, debit, credit, balance) must not drift into each other. Digital statements are read from the PDF's own text where it exists; scanned ones go through the same photo route as everything else in this guide, one sheet per page.

Which file types can you start from?

Almost anything that holds a picture of the table. A JPG from a camera is the title case, but the same engine takes PNG screenshots, scanned pages, and photos in the other common formats — the image to Excel converter and the format-specific pages are one engine wearing different doors, so a table that arrives as a screenshot converts through the same grid, the same number-typing and the same checks as a table photographed on paper. A table inside a PDF is the one case with its own route: if the PDF is a digital document, its text layer can be read directly, and if it is a scan inside a PDF wrapper, the statement converter's route — render the page, then read it — is the reliable path. Start from whichever file you actually have; converting formats first, by hand, is a step the tools make unnecessary.

What do you do with a heading that spans the whole table?

Real tables have furniture that grids don't: a title line across the top, a heading that spans three columns, a totals row in bold, a footnote underneath. Expect that furniture to arrive in the grid as ordinary rows and cells — a spanning heading typically lands as text sitting in one cell of its row rather than stretched across all of them. That is not corruption; it is the honest shape of the reading, and the grid is where you tidy it: delete the furniture rows you don't need, move the title out, or leave it exactly as printed if the spreadsheet is an archive of the page. The figures are what matter, and they are unaffected by where the title text sits. Check the totals row last and separately — it is both the easiest row to verify against the paper and the row that proves the columns above it landed straight.

What does the job look like from start to finish?

A worked example, with the numbers a real conversion produces. A supplier's printed price list, photographed on a desk: forty rows, four columns — item, pack size, rate, and a total column the supplier calculated by hand. The photograph goes in, the language is left on English, and the grid comes back with the forty rows standing in their columns. Two cells are flagged: a rate printed with a broken digit, and one item name in a stylised font. Both are fixed in the grid in seconds — the digit against the paper, the name by typing it. The totals column is summed in the downloaded spreadsheet and compared with the supplier's hand total: they agree, which closes the checking loop — the amounts were read as numbers, in the right columns, at the right values. The first conversion in a browser session takes a little longer than the rest, because the reading engine and the language data load once (a few megabytes, then cached); after that, each page is a matter of seconds. Forty rows that would have taken an hour to type — with the usual one-figure-in-fifty error rate — took one photograph and one careful check.

Is your table uploaded anywhere while this happens?

No. The reading, the rebuild and the spreadsheet generation all run in your browser, on your device. The image is never sent to a server — there is no account, no queue and no copy kept, because there is nowhere for a copy to be kept. A table of your own prices, your own accounts or your own customers is exactly the kind of document that should not need to visit a stranger's server in order to become a spreadsheet, and on this site it doesn't.

So the whole job is: photograph the table flat and level, pick its language, let the grid rebuild, check the flagged cells and the totals, and download a spreadsheet whose numbers are numbers. The typing marathon — and the one misread figure hiding in it — is optional now.

Originally published on CrushKB.