Extract a table from an image

Table extraction is not the same as OCR

Optical character recognition answers one question: what characters are in this picture? It reads left to right, top to bottom, and outputs a stream of text. For a paragraph that is exactly right. For a table it destroys the thing that made the data useful — which value belongs to which column.

Table extraction has to answer a second question: how is this content arranged? Where does the header row end, where does one column stop and the next begin, and which cells are empty rather than missing.

How GridFiller reads a table

1. 1. Find the tabular region

   The model locates the table within the image, ignoring surrounding navigation, titles, and marginal notes.

2. 2. Identify the header row

   Column names anchor the extraction and become the headers in your output.

3. 3. Map cells to columns

   Each value is assigned to a column, including where the boundary is implied by alignment rather than a drawn line.

4. 4. Return an editable grid

   The extracted table opens in an editor so you can correct anything before exporting.

5. 5. Export

   Download as CSV or XLSX, or keep the conversion in your history to revisit later.

Cases that need a little care

• Merged cells — a value spanning two columns has to be assigned somewhere, so check where it landed.
• Wrapped text — a cell whose content runs onto a second line can be read as an extra row.
• Repeated headers — page headers repeated mid-table are usually best removed in the editor.
• Footnote markers and asterisks attached to numbers.
• Columns of near-identical numbers, where a single misread digit is easy to miss.

Extracting one table from many images

When a table spans several images, upload them together and append rows: the header is taken once and the remaining rows are stacked in order. When the images come from different systems but describe the same thing, align by headers instead, so columns are matched by name rather than position.

If a result is not what you expected, you can re-run the same files with a different strategy without uploading them again.

Common questions

What if the image has more than one table?

Crop each table into its own image and upload them together with the keep-separate strategy, so each one lands on its own sheet.

Does it work on tables without gridlines?

Usually. Column boundaries implied by alignment are detected as well as ruled borders, though ruled tables tend to be more reliable.

How is this different from OCR?

Plain OCR returns characters in reading order and loses the grid. GridFiller reasons about layout, so the header row and column structure survive the extraction.