Find the earliest stage where information was lost
Retaking a weak source usually beats aggressive filtering. Diagnose the earliest point where information was lost, change one variable, and compare the same passage.
Use the diagnosis order to produce a fair before-and-after comparison, identify the change that reduced real errors, and avoid polishing an image without improving its text.
Test the suspected failure with OCR Quality Checker. Judge the controlled source change against How to Photograph a Document for Accurate OCR and Deskew, Dewarp and Perspective Correction for OCR.
The complete OCR accuracy guide separates capture, resolution, handwriting, and review problems into focused checks.
Diagnosis order
| Symptom | Test first | Avoid |
|---|---|---|
| Missing strokes | Sharper source | Upscaling as recovery |
| Lines merge or split | Geometry and segmentation | Aggressive sharpening |
| Fluent wrong words | Language and vocabulary | Judging by appearance |
| Only numbers fail | Field rules and character set | One global page score |
Run a controlled before-and-after test
Choose one passage that contains small print, punctuation, a proper name, and a number. Save the original OCR output and count exact corrections. If the image is blurred, recapture it. If lines tilt, correct geometry. If the source is clear but words are implausible, test language and segmentation. Change only one condition for each run.
Compare consequential errors rather than visual neatness. A darker image may look sharper while closing counters inside letters or deleting decimal points. Keep a change only when the same passage needs fewer meaningful corrections and no new failure appears elsewhere. This small experiment is more useful than repeatedly applying filters to the complete document.
- Preserve the untouched source
- Use the same passage and review rules
- Stop when the tested variable no longer improves real errors
Change one source variable and compare the same passage
Start with source pixels
Use the original scan or camera file. Upscaling creates more pixels but cannot reconstruct strokes removed by blur, clipping, or compression.
Correct geometry
Deskew rotation, flatten perspective, and keep characters a similar size across the page. Leave a small margin so edge text is not clipped.
Test layout and language
Choose a single block for prose and sparse modes for labels. Set the actual document language; similar glyphs and dictionaries affect recognition.
Measure the right sample
Use a representative page with small type, punctuation, names, numbers, and the hardest layout. Compare exact corrections rather than judging the output by appearance.
Controlled OCR accuracy test
- Save the original
- Choose a representative hard passage
- Count exact errors
- Change one variable
- Stop when that variable no longer reduces meaningful errors
Limits of an accuracy checklist
No preprocessing recipe repairs strokes that blur, glare, clipping, or compression removed. Improvements should be measured on the same representative text rather than judged from a cleaner-looking page.