How to Extract Tables from Scanned Documents (And Where It Fails)

Traditional optical character recognition (OCR) works in two isolated stages: first it detects character glyphs, and then a heuristic rule engine guesses where table borders exist based on lines or gaps.

Modern multimodal AI processes the entire image simultaneously. It reads the column headers (e.g., "Item", "Quantity", "Rate", "Amount") and uses semantic understanding to group values into the appropriate columns, even when border lines are absent or broken.

Even the strongest AI systems encounter physical limitations on scanned paper. Watch out for these four pitfalls:

• 1. Creased or folded paper: Folds cause text to bend across horizontal baselines, which can cause the AI to split one table row into two.
• 2. Low contrast and background bleed: Yellowed paper or carbon copy sheets where text bleeds through from the reverse side can introduce phantom characters.
• 3. Borderless multi-line item descriptions: When an item description wraps across three lines but the quantity and price are vertically centered, basic parsers align the numbers with the wrong line.
• 4. Skew and curvature: Photographing a thick bound book creates curvature near the spine, compressing table columns on the margin.

Follow these practical steps to extract tables from paper scans using Siftspindle:

1. Flatten the document: If photographing with a phone, take the picture directly from above in daylight with the camera parallel to the desk.

2. Sign in at siftspindle.com to access your 10 free monthly extractions.

3. Specify the table columns in your prompt: "Extract table with columns: Line Number, Description, Qty, Unit Price, Total".

4. Upload the scan (PDF, PNG, or JPG up to 25 MB).

5. Check the preview: Verify the row count against the original paper sheet, then export as an .xlsx Excel file.