Why Your Scan Quality Makes or Breaks AI Transcription (and How to Fix It)

Here's a scenario we see regularly: a user uploads a letter and gets a transcription with a few uncertain spots. They retake the photo with better lighting, upload again — and suddenly every word reads perfectly. Image quality is the single biggest factor you control in AI transcription accuracy. The AI model stays the same; what changes is the evidence you give it. This article breaks down the six most common image problems and shows you exactly how to fix each one.

How AI "Sees" Your Document

Unlike a human reader who can mentally fill in missing information, AI analyzes your image as raw pixel data. It identifies patterns — stroke shapes, curves, spacing — and matches them against learned letter models. When those patterns are clear, accuracy is remarkably high. When they're obscured by shadows, blur, or noise, the AI must estimate — and estimates are where errors creep in. The difference between "e" and "n" in Suetterlin is literally one additional arch stroke. At low resolution, that distinction vanishes.

Problem 1: Insufficient Resolution

What happens: Fine strokes merge together. The subtle differences between similar Kurrent letters ("e/n/m," "u/n," "d/f") disappear into fuzzy blobs. The AI can still produce text, but confidence drops sharply at ambiguous spots.

The fix: For scanners, set 300 DPI minimum — 400+ for faded text or pencil writing. For smartphones, ensure the text fills at least 80% of the frame and appears tack-sharp when you zoom in. If you can see individual pen strokes clearly on your phone screen, the resolution is good enough.

Problem 2: Shadows and Uneven Lighting

What happens: Shadows create bright and dark zones across the page. In dark zones, contrast between ink and paper drops — sometimes below the threshold where the AI can distinguish writing from background. Partial shadows (from your hand, a lamp, a fold in the paper) are worse than consistently dim lighting.

The fix: Overcast daylight near a window is ideal — naturally diffused, no hot spots. If using artificial light, place two lamps on opposite sides at equal angles. Flatbed scanners eliminate shadows entirely, which is why they produce the most consistent results. If you see any shadow on the document when framing your shot, reposition before shooting.

Problem 3: Perspective Distortion

What happens: Photographing at an angle stretches letters at the far edge and compresses them at the near edge. Lines that should be horizontal become angled. The AI can tolerate minor tilt (5–10°), but steeper angles measurably reduce accuracy, especially at page margins.

The fix: Hold your phone directly overhead, parallel to the document surface. Rest your elbows on the table for stability. Scanner apps (Adobe Scan, Microsoft Lens) detect and auto-correct mild perspective distortion — use them as your default camera for documents.

Problem 4: Grid Lines and Ruled Paper

What happens: Diaries and notebooks from the early 20th century often used heavily ruled or grid paper. When the grid lines are dark relative to the ink, the AI can misinterpret grid segments as letter strokes — particularly thin horizontal strokes that resemble parts of Suetterlin characters.

The fix: Maximize ink-to-background contrast. Use "document mode" in scanner apps, which brightens the background and sharpens the writing. On a flatbed scanner, increasing brightness slightly while maintaining contrast can push grid lines closer to white while preserving ink visibility.

Problem 5: Faded Ink on Yellowed Paper

What happens: Century-old documents often have ink that has faded from black to light brown — on paper that has yellowed to a similar shade. The contrast ratio drops from maybe 90% (black ink on white paper) to 20% or less. At this point, the AI struggles to separate writing from background.

The fix: Always scan in full color, never grayscale or black-and-white. Color mode preserves spectral differences between ink and paper that are invisible in grayscale — aged iron-gall ink and yellowed paper often have different hues even when they have similar brightness. A gentle contrast increase in any image editor (+10–20%) can also help, but don't overdo it or you'll introduce artifacts.

Problem 6: Over-Compressed JPEGs

What happens: Aggressive JPEG compression introduces "ringing" artifacts around letter edges — faint halos and blocks that smear fine details. Since Suetterlin's letter distinctions depend on tiny features (a single arch stroke, a small mark above "u"), compression artifacts can erase exactly the details that matter most.

The fix: Save as PNG (lossless, larger files) for archival quality, or JPEG at 90%+ quality (good balance of size and fidelity). Never re-save a JPEG — each save cycle degrades quality further. If you're uploading someone else's photos, ask for the original files rather than versions that were shared via messaging apps (WhatsApp, iMessage, etc. often compress images aggressively).

Quick Reference Checklist

Before uploading to GermanLetters, verify: ✓ Text is sharp when zoomed in. ✓ No shadows visible anywhere on the page. ✓ Page is straight and fills 80%+ of the frame. ✓ Scanned in color at 300+ DPI (or crisp phone photo). ✓ Saved as PNG or high-quality JPEG. ✓ No visible compression artifacts around letters.

GermanLetters's AI handles imperfect images — creased pages, foxing, bleed-through, and moderate fading are all manageable. But the closer you get to ideal conditions, the better your results will be. For detailed photography and scanning techniques, see our complete digitization guide. Ready to test with your documents? Upload your first page free at GermanLetters.