Retake missing evidence before trying to filter it back
A useful document photo is square to the page, sharp at the smallest important characters, evenly lit, and complete enough to preserve page context.
The procedure ends with both an untouched full-page source and a recognition crop whose smallest important text remains sharp, evenly lit, complete, and large enough to inspect.
Put the procedure into practice with JPG to Text. Continue with Deskew, Dewarp and Perspective Correction for OCR and Cropping Borders and Background Noise Before OCR.
For the steps after capture, see the OCR image preparation collection covering perspective, borders, grayscale, and contrast.
Camera check before recognition
- Clean the lens
- Disable beauty, portrait, or background-blur modes
- Check all four corners for focus
- Keep page numbers and units
- Compare the OCR output with the original photo
Build a sharper, flatter, evenly lit source
Make the page and sensor parallel
Place the page on a contrasting surface and hold the phone directly above its center. Check that opposite page edges are close to parallel. Perspective correction can reshape a trapezoid, but it cannot restore detail from a far corner that was captured too small or out of focus.
Focus on the smallest consequential text
Tap near small dates, footnotes, serial numbers, or decimal values rather than a large heading. Review the captured image at full size. If a narrow stroke blends into the background, the OCR engine receives the same ambiguity.
Move reflections instead of masking them
Use broad, indirect light from both sides when possible. Shift the lights or camera until glossy glare moves away from text. A contrast filter can darken surviving strokes, but a white reflection contains no recoverable ink detail.
Keep a full-page source and a working crop
Save the uncropped original for context, then make a derivative crop for recognition. The crop should remove unrelated surroundings while retaining every line, page number, label, and unit needed to interpret the text.
Use the smallest field as the capture test
A delivery note may have a large supplier name but tiny quantities and a faint reference code. Frame the entire page, keep the sensor parallel, and focus near the smallest consequential field rather than the headline. Inspect all four corners at full size before accepting the photo.
If glare crosses one total or the far corner is soft, retake the page. Keep that uncropped image for context, then make a working crop that removes the desk while preserving page edges, labels, units, and identifiers. The recognition result should be checked against the original capture, not against a filtered derivative alone.
- Retake missing detail
- Keep full-page context
- Inspect corners, glare, and small fields
Capture defect and best next action
| What you see | What OCR loses | Best next action |
|---|---|---|
| Soft edges everywhere | Character stroke boundaries | Retake after cleaning the lens and stabilizing the phone |
| One corner is smaller | Detail and consistent character scale | Reposition the camera parallel to the page |
| White patch across text | All ink under the reflection | Move the light or camera and retake |
| Dark fold or gutter | Contrast and complete glyphs | Flatten gently or capture each page separately |
A phone image may still be insufficient
Good technique reduces avoidable loss but cannot overcome a damaged original, text below the camera's resolved detail, reflective material, or a page that cannot be flattened safely.