Image to Excel, worked out on your own device
Extract a table from an image
Paste a screenshot or drop a photo of a table. Rows, columns and merged cells are detected on your device and exported to Excel — nothing is uploaded.
Table extractor
Picture
Ctrl+V
Paste a screenshot of a table
or drop an image here
Drag a box on the picture to extract only that part.
Table
| A | B | C | D | E | F | |
|---|---|---|---|---|---|---|
| 1 | ||||||
| 2 | ||||||
| 3 | ||||||
| 4 | ||||||
| 5 | ||||||
| 6 |
Cells land here, with merged cells kept merged.
Yellow cells were hard to read. Click any cell to correct it before you export; edits go into every download.
From screenshot to spreadsheet in three moves
- Paste or drop the table. A crisp screenshot works best — the Snipping Tool (Win+Shift+S) or ⌘+Shift+4 on a Mac copies straight to the clipboard, so a single paste is enough.
- Check the yellow cells. Each cell's text comes with a confidence score; anything under 85% is shaded so you can fix it in place. Blue, orange and purple frames on the picture show which rows, columns and merged cells were found.
- Export. The .xlsx keeps merged cells, bolds the header rows and stores numbers like 1,152 or 13.1% as real numbers. CSV and the clipboard copy keep the text exactly as read.
What it reads well, and what trips it up
Measured on this page's own test set with the built-in models (Apple M2 Max; WebGPU about 0.4 s per table; the CPU path, which is single-threaded WebAssembly, about 3.7 s):
Reads well
- Screenshots of web tables, spreadsheets and reports — 14 × 5 borderless finance table: 70 of 70 cells right.
- Header rows with merged cells (colspan and rowspan) — all four merged cells of the sample found.
- A table inside a page: the title and the note above it, and a footnote below, are left out of the grid.
- A slightly tilted, blurry phone shot of a printed table: 40 of 44 cells right (“Cherry” read as “Chemy”), which is why yellow cells exist.
Gets it wrong
- Dense tables photographed at a steep angle: an old bus timetable photo came back with only 9 of its 20+ rows.
- Warped or curved pages — straighten them first with Flatten document photo.
- Handwriting, stamps and checkbox ticks: the text reader is trained on printed type.
- Two-column prose that only looks like a table. If you just need the words, Image to text keeps paragraphs intact.
How it works: Microsoft's Table Transformer v1.1 (MIT) finds the table, its rows, columns, header and merged cells; PaddleOCR PP-OCRv5 (Apache-2.0) finds and reads each line of text; each line then goes into the row × column cell its centre falls in. All three models run in this tab through ONNX Runtime Web.
Questions
Is my image uploaded anywhere?
No. The picture is decoded in your browser, the three models run in this tab (WebGPU, or WebAssembly on the CPU), and the Excel file is written by JavaScript on your machine. The only downloads are the model files themselves, once. You can disconnect after the first table loads and keep converting in the same tab.
Does it keep merged cells?
Yes, as far as the structure model detects them. Table Transformer marks spanning cells, and the grid merges every row × column position whose centre falls inside one. In the .xlsx they become real merged ranges; in CSV the text sits in the top-left cell and the others stay empty, which is how CSV has to do it.
Can it read handwritten tables?
Not reliably. Rows and columns of a hand-ruled table are often found, but the text reader (PP-OCRv5 mobile) was trained on printed and on-screen type, so handwritten entries come back garbled or empty. Typed forms, printouts and screenshots are what it is built for.
Why did some text land in the wrong cell?
Almost always because the picture is skewed or the text is tiny. Each text line is placed by its centre point, so a tilt of a few degrees across a wide table can push a line into the next row. Take the screenshot straight on, crop to the table (drag a box on the picture), and make sure the smallest text is at least about 10 pixels tall. For photos of paper, flatten the page first.
Can I convert several images or a whole PDF at once?
Not in the free page: it takes one picture and extracts one table (the largest, if there are several — drag a box to pick another). Batches of images, PDF pages and several tables per page are planned for a Pro tier. If what you really have is a folder of invoices or contracts with a text layer, AIFormFerry, our Mac app, turns the fields you name into one spreadsheet row per document.
How is this different from iLovePDF, Nanonets or Excel's Data from Picture?
Those services send your image to a server to be read (Excel's Data from Picture also needs Microsoft 365). This page does the whole job on your computer, has no daily limit and no sign-up, and shows you where the structure was detected so you can see why a cell went wrong. The trade-off: server-side models are larger and handle messy photos better.
Pictures, cell text and exports never leave this browser tab. Models are cached by the browser so the second table starts in about a second.