OCR notes
OCR is often blamed for scanner faults and source damage. The recognizer sees only the bitmap it receives. A better scan of a bad page can beat a perfect engine fed with a crushed one-bit image.
Failures I see repeatedly
| Bad input | Common OCR result | What helps |
|---|---|---|
| Low contrast carbon copy | Missing small words, numbers read as spaces | Grayscale scan, manual levels, no aggressive thresholding. |
| Skewed page | Line breaks in the wrong place | Deskew gently. Avoid using a black border as the page edge. |
| Handwritten annotations | Marginal notes enter the main text | Crop to the printed body, keep original image for context. |
| Dot-matrix print | m/r/n confusion and broken punctuation | Scan at 300 or 400 DPI grayscale. Avoid JPEG compression. |
| Photocopy generation loss | o/e/a collapse together | Try unsharp mask lightly, then proofread against the image. |
My plain rule
Archive the image first and the OCR second. The text can be corrected later. The scan cannot be recovered if the first capture threw away the detail.
For more examples of damaged pages and OCR edge cases, see the outside resources on the links page.