Flatbed Notes

OCR trouble list

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 inputCommon OCR resultWhat helps
Low contrast carbon copyMissing small words, numbers read as spacesGrayscale scan, manual levels, no aggressive thresholding.
Skewed pageLine breaks in the wrong placeDeskew gently. Avoid using a black border as the page edge.
Handwritten annotationsMarginal notes enter the main textCrop to the printed body, keep original image for context.
Dot-matrix printm/r/n confusion and broken punctuationScan at 300 or 400 DPI grayscale. Avoid JPEG compression.
Photocopy generation losso/e/a collapse togetherTry 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.