How to Extract Tables from Images and PDFs into Excel or CSV

Before picking a tool, find out what kind of table you actually have, because the three common sources fail in different ways. Open the PDF and try to select a single number with the cursor. If the number highlights on its own, the file is a text PDF: every character is stored as text at an exact position, and extraction is mostly about grouping those characters into rows and columns. If the whole page highlights as one block, or nothing highlights at all, you are holding a picture of a table, and someone has to read the pixels first.

The third case is easy to miss. Many reports exported from older accounting systems, terminal programs and email bodies look like tables but are plain text lined up with tabs or runs of spaces. There are no cells at all, only a monospaced layout. Running OCR on a screenshot of that text throws away information you already have; copying the text and splitting it into columns is faster and exact. The cards below sum up what each source needs.

Table recognition is a second pass that runs after the characters are known. First it finds cells: on ruled tables it traces the horizontal and vertical lines and treats each enclosed box as a cell; on borderless tables it looks for vertical gutters of empty space that run the full height of the block. Then it aligns columns, deciding that a number sitting a few pixels to the right of the one above still belongs to the same column.

Next comes the header row. Recognizers look for a first row that is bold, shaded, or made of words above columns of numbers, and they join headers that wrap over two lines into a single cell. Merged cells are the last step: a label such as Q1 spanning three month columns is detected from the missing internal lines and either rebuilt as a merged cell or copied into each column it covers.

None of this is reading comprehension. The software does not know that a column holds dates or that a row is a subtotal; it only sees geometry and characters. That is why the same tool can be flawless on one statement and scramble the next one from a different bank, and why settings such as Table style and Detect header row exist at all: they hand the recognizer a fact it cannot reliably infer.

Most failures fall into two groups. Geometry failures break the grid, so values land in the wrong row or column. Number failures leave the grid intact but change what a cell means, which is worse because the sheet looks fine. A misplaced column is obvious at a glance; 1.234 read as a decimal instead of a thousand is not.

Locale is the main source of number trouble. The United States and the United Kingdom write 1,234.56; Germany, Brazil and most of continental Europe write 1.234,56; France uses a space as the thousands separator; Switzerland often uses an apostrophe, as in 1'234.56; and India groups digits as 1,23,456. Accounting documents add negative numbers in parentheses and currency symbols glued to the digits. Here is what to watch for and what fixes each case.

Every minute spent on the input saves several in the spreadsheet. OCR engines are commonly tuned for text scanned at around 300 dpi, which for an A4 or US Letter page means roughly 2480 × 3508 px or 2550 × 3300 px. A modern phone camera easily beats that if the page fills the frame, but a full-page photo shot from across the desk, then cropped to a small table, can drop well below it.

Contrast matters as much as resolution. Thermal-paper receipts fade, gray table rules on a gray background disappear, and a shadow from your phone across the page can cut a column in half. If the original is a paper document you will extract from regularly, the guide to scanning documents to PDF covers lighting and scanner settings in more detail. Before you upload, run through this list.

Excel's Data from Picture arrived in the Android app in 2019, then on iPhone, and it is now in Excel for Microsoft 365 on Windows and Mac under Data > From Picture. You photograph or load one table, review the recognized cells in a side panel that flags doubtful values, and insert them. Google offers a less direct path: in Drive, right-click an image or PDF and choose Open with > Google Docs. Google's own help page limits this to files of 2 MB or less and reads only the first 10 pages of a PDF, and the result is a document, so the table often arrives as lines of text you then paste into Sheets and split.

Adobe Acrobat Pro exports a PDF to a Microsoft Excel workbook and runs OCR on scanned pages during export. Excel for Microsoft 365 on Windows can also pull tables from a text PDF with Get Data > From File > From PDF, which opens Power Query; it lists every detected table by page, lets you append a statement's monthly pages into one table and refreshes when you swap in next month's file. It does not read scans, and Excel for Mac does not have the connector.

Browser converters cover the rest without installing anything. For a text or scanned PDF, PDF to Excel turns every detected table into its own sheet; for a photo or screenshot, convert the image to Excel or save the table as CSV when another program will import it. The table compares the five routes.

A converted table that looks right is not yet a table you can rely on. Start with the row count: if the statement lists 47 transactions, the sheet should have 47 data rows, not 46 because two lines merged or 49 because a wrapped description became its own row. Then add up one numeric column with SUM and compare it with the total printed on the document. If the document has no total, add a few values by hand and compare.

Next, look for cells that only pretend to be numbers. In Excel, real numbers align right by default and text aligns left, so a column with a few left-aligned amounts is telling you something; ISNUMBER on a helper column finds them all at once. Dates deserve the same suspicion. A date read as text sorts alphabetically, so 10/02 lands before 2/03, and 03/04 can mean 3 April or March 4 depending on who wrote it.

Finally, check what was lost rather than what was changed. Long account and card numbers are a classic trap: Excel stores numbers with 15 significant digits, so a 16-digit card number typed or imported as a number has its last digit turned into a zero.

Pick XLSX when a person will work with the data in Excel, Google Sheets or Numbers. It keeps several tables on separate sheets, real number and date types, bold header rows and merged cells. Pick CSV when software will import it: budgeting apps, accounting packages, databases and scripts. CSV is plain text with one table per file, described in RFC 4180 from 2005, and the three settings that matter are the delimiter, the encoding and whether numbers are stored as numbers.

The delimiter has to match the importing program; Excel in countries that write decimals with a comma expects semicolons. The encoding decides whether names such as Müller or São Paulo survive: UTF-8 with BOM is what Excel for Windows reads correctly on a double-click. Leading zeros are the classic loss. US ZIP codes such as 02134, employee IDs and product codes are identifiers, not quantities, and become 2134 as soon as anything treats them as numbers. Turn off Store numbers as numbers for tables full of IDs, or import the CSV with those columns set to Text.

Extracting one table is a lookup; extracting a year of invoices or twelve monthly bank statements is a small data project. Keep the source documents in the same layout where you can, and convert them in one pass so every file gets the same settings. With text PDFs, upload up to 20 at a time, each up to 95 MB, and choose Everything on one sheet in the Sheets option when a statement's transaction table continues across pages. Photos convert one table image at a time, so for a stack of paper invoices it is often quicker to combine the pages with image to PDF first and point PDF pages at the range that holds the tables.

After a batch, append the results into one sheet with a column that records the source file name. If the documents are mostly text with a table or two, converting the PDF to Word keeps the paragraphs and the tables together instead.

Bank statements, payslips and invoices carry account numbers, addresses and spending history, so think about where they go. On UltKit, uploads and the generated files are deleted 30 minutes after processing and are not read by anyone. Still, the safest copy is the one that never leaves your computer: crop or redact columns you do not need, such as full account numbers, before upload, and delete local exports you no longer need.