Extract a Table From an Image
Turn a screenshot, scan, or photograph of a printed table into reviewable rows and a CSV download. OCR runs with Tesseract, then nearby word boxes are grouped into rows and cells. This is a useful starting point for receipts, schedules, and simple reports, but it is not a perfect grid reconstruction: merged cells, borders, formulas, rotated text, and irregular columns still need checking.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeWhat it does
- Runs OCR locally in the browser
- Groups word positions into inferred rows and cells
- Shows confidence and a table preview
- Copies TSV or downloads CSV
- Warns when layout cannot be reconstructed exactly
How to use OCR Table Extractor
- 1
Choose the table image
Upload a screenshot, scan, or straight photograph of a printed table. Clear images with even lighting give the most reliable word positions.
- 2
Start the recognition
Press Extract table. Recognition takes a few seconds on a full page, and longer on a photograph than on a screenshot.
- 3
Review the inferred rows
Check the preview, especially numeric cells, column boundaries, merged headers, and rows with wrapped text. The confidence score is a recognition signal, not a proof of correctness.
- 4
Copy or download
Copy tab-separated values for a spreadsheet or download a CSV copy, then compare important values with the source image before using them.
How it works
The browser sends the image to a local Tesseract worker. Tesseract returns recognized words, confidence scores, and bounding boxes rather than only one long string. The table extractor sorts those words by their vertical position, groups nearby baselines into rows, and uses larger horizontal gaps to separate cells. The result is shown before download so the inferred structure can be corrected or rejected by a human.
This approach is deliberately modest. A table is not just text; it is a grid with relationships between cells. A line, merged header, wrapped label, or blank cell can change the meaning even when every character is recognized correctly. The tool therefore produces a CSV starting point and describes the geometry heuristic instead of pretending to recreate an Excel workbook exactly.
Improve the result
Use a straight screenshot when possible. For photographs, place the camera parallel to the page, use even light, avoid glare, and crop away unrelated background. Make sure the smallest numbers are large enough to read at normal zoom. If the preview combines two columns, add more space between columns in the source image or use a crop of the table. If rows drift vertically, correct the photograph rather than trusting the CSV.
Choose an image for table extraction
Review before use
The best source is a straight, tightly cropped screenshot or scan with consistent spacing between columns. Perspective, glare, shadows, merged headers, and wrapped descriptions make the geometry harder to infer even when characters are recognized correctly. For a photographed page, hold the camera parallel to the paper and use enough light to avoid motion blur. Crop unrelated material before loading the image, but do not crop away column headings that explain the values below them. After export, retain the original image and check important rows before sorting or filtering the CSV.
Compare headers, decimal points, minus signs, dates, account numbers, and totals with the source. Check rows that contain wrapped text or blank cells because they are the most likely to shift. The output is a local extraction aid, not an audit record. Keep the original image and treat the CSV as untrusted until the important values have been verified.
Examples
Convert a schedule screenshot
Straight screenshots usually preserve spacing well enough for a quick starting table.
Extract a photographed receipt table
Re-photograph the receipt straight-on and inspect every amount; OCR can recognize a digit incorrectly even when the row layout looks right.
Frequently asked questions
Does this create a real Excel spreadsheet?
It downloads CSV, which opens in Excel, Google Sheets, and other spreadsheet programs. It does not recover formulas, formatting, merged cells, hidden rows, or the original workbook structure, so review the result before treating it as a finished spreadsheet.
How are columns detected?
The browser groups words with nearby vertical positions into rows, then starts a new cell when the horizontal gap is large enough. This layout heuristic works best for simple tables with consistent spacing and can misread complex grids.
Is the image uploaded?
Yes. The image is sent to our server, recognized with Tesseract there, and the rows are returned. The upload is deleted as soon as recognition finishes and the extracted table is not stored at all. There is no account and no history.
Can it read handwritten tables?
Not reliably. The engine is trained for printed text, and handwriting, faint ink, glare, blur, perspective, and decorative fonts can all produce incorrect cells. Proofread important values against the original image, especially names and numbers.
Why is the confidence score not an accuracy guarantee?
Confidence describes how certain the OCR engine was about recognized characters. It does not prove that the right column was inferred or that a similar-looking digit is correct, so review layout and values separately before exporting.
Further reading
- The Same Table Was Three Times Bigger as JSONCSV writes each column name once, at the top. JSON writes it again on every single row. On a table of any size that is not a subtlety — it is the dominant term, and it explains both why the conversion inflates and why it barely matters over the wire.
- JPEG Compression Made No Difference to OCR Accuracy At AllScanning guides tell you to save as TIFF or PNG because JPEG artifacts confuse OCR. It sounds obviously right — lossy compression smears exactly the hard edges that letter recognition depends on. We measured it across a wide range of quality settings and found no difference whatsoever, which changes the storage arithmetic for anyone digitising a lot of paper.
- OCR Reads Nothing at 72 DPI and Everything at 150The standard advice for scanning documents you intend to run through OCR is 300 DPI. We measured where the accuracy actually changes, and it is not a gentle curve with 300 comfortably on the good end. It is a cliff between 96 and 150, and everything above 150 was identical.