Turn Table Images into Excel — Rebuild the Grid, Not Just Read the Text
Manually typing a printed table into Excel takes 3 minutes per page — and when merged cells and borderless grids confuse the OCR, fixing the output takes longer than typing it from scratch. The real challenge isn't reading text from a picture; it's rebuilding the table structure — which cell belongs to which column, which row aligns with which header. Define the columns you want, and the visual AI reconstructs the grid from pixel images by understanding cell relationships, not just cell contents.
5-10s per page · 99% accuracy on printed tables · Handles merged cells, borderless grids & nested tables
What You Can Extract from Any Table Image
Point a phone at a printed grid, take a screenshot of a web dashboard, or scan a spreadsheet — you're not just converting an image to text. You're reconstructing a table. That's where Custom Column Extraction changes the equation: instead of dumping all detected text into a spreadsheet and hoping the cells land in the right place, you type the column names you want — Invoice Number, Vendor, Total — and the AI finds those values on every image by understanding what they mean and where they sit in the grid structure. The column names you type become the exact headers in your output.
These are examples of column names you type. Any column name works — the AI finds matching values by understanding the grid structure, not by template matching. The output is one clean spreadsheet with exactly your columns as headers.
Image-to-Excel Is a Table Reconstruction Problem — Not a Text Reading Problem
When you point a phone at a printed table, take a screenshot of a web dashboard, or scan a price list, you're not just converting "image to text." You're converting a visual layout of pixels into a navigable spreadsheet where every cell maintains its relationship to its column, its row, and its neighbors. The software that misunderstands this distinction gives you a text dump that needs hours of manual re-alignment. Here's the anatomy of that failure — and what happens when the approach matches the problem.
Why Traditional OCR and Grid Parsers Break on Images of Tables
OCR reads cells; it doesn't read cell relationships. Standard OCR — including what powers Excel's built-in "Data from Picture" and most online converters — works pixel by pixel, detecting character shapes and outputting text strings with approximate X,Y positions. It can tell you that "$1,290.00" exists at coordinates (450, 320), but it cannot tell you that this is the Total row of the table. The relationship between a header saying "Total" and the number below it is invisible to position-based tools — they extract two unrelated text fragments and dump them into adjacent spreadsheet cells, hoping proximity equals belonging. When the table has no grid lines separating columns, or when headers span merged cells, the output scrambles.
Merged cells, borderless tables, and nested grids are invisible to grid detection. Grid-based parsers — tools that look for lines and use them to split rows and columns — fail the moment the lines aren't there. A price list table where columns are separated by whitespace alone, not drawn borders, produces no grid to detect. A dashboard screenshot where "Q1 Revenue" spans four sub-columns with a merged header gets flattened into a single-cell label, assigning the data underneath to the wrong columns. When tables are nested — a summary table containing an embedded detail table — the outer and inner grids collide, and the parser can't determine which structure takes priority.
Fixed templates break the moment the table layout changes. Template-based tools (Nanonets, Rossum) require you to configure extraction rules tied to specific document layouts. When invoice #43 has the Total in the bottom-right corner and invoice #44 has it centered above a "Payment Due" section, the template misses. The same applies to screenshots from different apps, scanned price lists with different column arrangements, and phone photos of tables taken from slightly different angles. Every layout variation is a new failure point — and you're the one fixing it.
How Column-Name Extraction Rebuilds Grid Structure by Understanding Relationships
You define what matters — the AI maps it across the grid. Type "Invoice Number, Date, Vendor, Subtotal, Tax, Total" — the visual language model reads the entire image as a table structure, understanding that "Invoice Number" refers to a value like "INV-2024-0842" sitting near the top header area, "Total" refers to the bottom-line amount, and "Vendor" is the company name printed near the billing section. It doesn't need the columns to be labeled with those exact words — it finds values by understanding their role in the table's layout. This is Custom Column Extraction: instead of accepting whatever text the OCR finds dumped into whatever cells it guesses, you name the columns and the AI reconstructs the grid around your target fields.
Merged cells, borderless tables, and nested grids are understood in context. The visual large model sees the full layout — it recognizes that a header spanning Q1 through Q4 is a parent grouping, not a single-cell label. Borderless tables separated by whitespace are read as columns because the AI understands visual alignment, not because it detected lines. When a screenshot of a dashboard contains a summary table with an embedded detail table, the AI distinguishes the two structures by understanding layout hierarchy — the same way a person looking at the image would. Merged cells retain their relationship to their sub-columns; data under each sub-column is assigned correctly.
One column definition works across any image — regardless of layout. You have a screenshot of a web dashboard, a phone photo of a printed supplier price list, and a scan of a spreadsheet from 2018 — each with different column positions, different header wordings, different numbers of total columns. Upload them all as one batch. Define your column names once. The AI reads each image independently, reconstructing the grid structure for each one, and merges the results into one spreadsheet. Processing takes 5-10 seconds per image. The same approach handles multi-page tables — the AI recognizes that a table continues across page breaks, preserves row continuity, and doesn't duplicate repeated headers as data rows. Use Collection Link — a shareable upload page — to let others submit table images directly to your processing queue without needing an account.
Your Table Is Trapped in a Picture — Here's How It Gets Out
You didn't choose to receive a screenshot instead of a spreadsheet. You didn't ask your vendor to send a photo of their price list instead of a CSV. But here you are, staring at a picture of a table, needing the data in Excel. Here's the path from pixel grid to spreadsheet grid.
Upload the Image — Phone Photo, Screenshot, or Scan
The image can be a screenshot of a web dashboard, a phone photo of a printed vendor price list, a scanned spreadsheet from years ago, or even a picture of a whiteboard grid from a meeting. The tool accepts JPG, PNG, WebP, and PDF. You can upload multiple images at once — different sources, different formats, different table layouts — and batch-process them together.
Type Your Column Names — Once for All Images
Type the columns you want: Invoice Number, Date, Vendor, Description, Quantity, Unit Price, Total. The AI reads each image with these targets in mind, reconstructing the table grid around the values that matter — not dumping every text fragment it finds. If you need calculations built into the output, use Computed Columns: write a column like Line Total (Qty × Unit Price) and the AI performs the math during extraction. For classification, Inferred Columns let you name a column like Category (options: Raw Materials/Finished Goods/Office) — the AI determines the correct category by reading the table content.
Download One Merged Excel File
Processing takes 5-10 seconds per image — roughly 18x faster than manual entry. Each image becomes one or more rows. Each column is exactly what you named. The dashboard screenshot, the vendor price list photo, and the scanned spreadsheet all land in one XLSX file with matching headers. Export as XLSX, CSV, or JSON. If a value isn't readable on a specific image — extreme glare on a phone photo, for instance — the cell stays empty rather than being filled with a hallucinated value.
When Table Reconstruction Works — and When Input Quality Matters More Than AI Smarts
AI-powered grid reconstruction handles structural problems that break position-based tools — merged cells, borderless layouts, variable column counts. But it works with what's visible. Understanding where quality matters, and where it doesn't, sets realistic expectations.
When It Works Best
Screenshots of web dashboards, data tables, and report grids. Screenshots are the most reliable input — no perspective distortion, no lighting issues, clear text at screen resolution. Up to 99% accuracy on printed, clearly rendered table data.
Tables with merged cells, spanning headers, and borderless layouts. The AI reads structure semantically — it recognizes that a header spanning multiple sub-columns is a parent label, and that columns separated by whitespace are still columns. These are failure points for position-based OCR; they work here because layout is understood visually.
Batch processing across different table layouts and sources. One screenshot from an ERP system, one phone photo of a supplier price list, one scanned spreadsheet — upload them together with one column definition and get one merged Excel output.
Multi-page tables that span several image files. The AI recognizes table continuity across pages, maintains row numbering, and identifies repeated headers — avoiding the fragmented, duplicated-header output that page-by-page parsers produce.
When to Be Cautious
Phone photos with extreme glare or heavy shadows. A bright glare spot directly over a critical value, or a shadow that completely occludes text, will prevent extraction on that field. The AI reads what's visible — it cannot reconstruct data hidden by reflection. Angle the phone to avoid direct light on the document surface.
Nested sub-tables within a larger grid. A table containing an embedded sub-table — a summary section with its own mini-grid of detail rows, for instance — may cause the outer and inner structures to collide. The AI attempts to merge them into a coherent flat table, but complex nesting can produce misalignment. Extract the main table and sub-table separately for best results.
Hand-drawn tables and hand-annotated grids. Printed tables work best. Tables drawn by hand with irregular lines, inconsistent column widths, and handwriting introduce ambiguity in both structure detection and text recognition. Neat printed handwriting on a clearly lined grid works reasonably; freehand sketches and faint pencil marks will produce lower accuracy.
Low-resolution images below 150 DPI or heavily compressed screenshots. Severely pixelated or JPEG-compressed images where characters blur together reduce text recognition accuracy. The AI compensates with contextual understanding, but below a threshold — roughly the point where a person squinting at the screen can't read the numbers either — extraction will be unreliable. A clear image is always the best input.
Frequently Asked Questions
Does AI image-to-excel extraction handle merged cells and borderless tables — or does it need visible grid lines?
It handles both. The visual language model reads table structure by understanding visual groupings — the way a person looks at a table and recognizes where columns start and end without tracing grid lines. Borderless tables where columns are separated only by whitespace are read correctly because the AI understands column alignment, not line detection. Merged cells spanning multiple sub-columns are recognized as parent labels, not flattened into a single-cell value. Position-based tools like Tabula and Camelot fail on borderless tables entirely and misalign merged-cell data — this is the fundamental difference between grid-line detection and semantic structure understanding.
Can I extract only specific columns — like Invoice Number, Date, and Total — or does it pull everything visible in the table image?
You control the columns entirely. Type the field names you want — Invoice Number, Date, Vendor, Total, Tax — and the AI extracts only those values. Everything else in the image is ignored. This is the core of Custom Column Extraction: you define the output schema, and the AI maps data into it by understanding what each field means in the context of the table. If you don't specify columns, the AI auto-detects all visible fields — useful as a quick starting point. You can also use Computed Columns to build calculations into the extraction (e.g., Tax Amount (Subtotal × 0.08)) or Inferred Columns to classify rows (e.g., Category (options: Food/Beverage/Office Supplies)) based on table content — the column names you type become your Excel headers.
How accurate is table extraction from phone photos compared to screenshots or scanned documents?
Screenshots of digital tables produce the most accurate results — up to 99% for clearly rendered text at screen resolution. Phone photos introduce variables: lighting angle, shadows, perspective distortion, and camera resolution. The AI compensates for moderate variations by reading context — it understands that a number next to "Total" on a slightly angled receipt photo is still the Total. Severe cases — extreme glare washing out data, motion blur, or heavy shadows fully blocking characters — will reduce accuracy on the affected fields. A practical rule of thumb: if you can read the value on the image, the AI can too. Processing takes 5-10 seconds per page, roughly 18x faster than manual entry (~3 min manual per page vs ~5-10s automated).
Can I process multiple table images at once — different layouts, different sources — into one spreadsheet?
Yes. Upload multiple images in one batch — screenshots, phone photos, scanned spreadsheets, mixed formats — and define one set of column names. Each image is processed independently against the same column targets. A screenshot of a web dashboard, a photo of a printed vendor price list, and a scan of an old invoice all produce rows in the same output spreadsheet with matching headers. For ongoing workflows where you receive table images from other people — vendors sending price lists, field staff photographing inspection grids, clients submitting forms — use Collection Link: generate a shareable upload URL, send it to anyone, and they upload images directly through a browser to your processing queue. No account or app required on their end.
What about images of tables that span multiple pages — like a scanned price list that's 8 pages long?
Upload all pages as separate images and batch-process them with one column definition. The AI reads each page and outputs rows in order. Multi-page PDFs are also supported — the tool processes them as a single document and recognizes table continuity across page breaks, preserving row order and identifying repeated header rows so they aren't duplicated as data. Position-based tools treat each page in isolation, producing separate outputs that require manual merging and header-stripping — a known frustration for anyone extracting data from a scanned catalog, rate sheet, or report where tables run across dozens of pages.