OCR Accuracy: Fix the Scan Before Blaming the Engine
OCR errors usually start before the OCR engine sees a single character. A tilted phone photo, a shadow through the middle of a receipt or tiny blurred text can destroy recognition long before “accuracy settings” enter the picture.
The fastest improvement is often boring: fix the source.
Straight text beats a giant image
A moderate-resolution scan with level text lines is easier to recognise than an enormous photo taken at an angle. Perspective distortion changes letter shapes across the page. Motion blur wipes out edges. Neither problem is repaired simply by giving the image more pixels.
If you are photographing a document, hold the camera parallel to the page, fill the frame without cutting off corners, and use steady light. A dark page with a bright window reflection is a harder OCR problem than a smaller but evenly lit scan.
Contrast does more work than people expect
Grey text on a grey receipt is difficult. So is a book page with a heavy gutter shadow. Crop away the desk, straighten the page and make the foreground clearly different from the background. Avoid over-sharpening; halos around characters can create new shapes for OCR to misread.
Some characters deserve automatic suspicion
- 0 and O in reference numbers.
- 1, l and I in codes, email addresses and names.
- Decimal points and commas in financial figures.
- Hyphens, apostrophes and small punctuation.
- Dates where one wrong digit changes the meaning.
Names, account references, invoice totals and dates are exactly the fields people tend to copy without rereading. Check them first.
A searchable PDF can look perfect while the OCR is wrong
A searchable PDF often keeps the original scan visible and places recognised text invisibly underneath it. The page can look identical to the source even when that hidden text layer contains mistakes. Visual inspection alone will not catch those errors.
After export, search for two or three words you know are present. Copy a sentence into a plain-text editor. If the file will feed another system, test a real extraction before processing a whole archive.
Tables, forms and handwriting are separate problems
Plain printed paragraphs are the easiest case. Tables add column relationships. Forms mix labels, boxes and handwritten values. Handwriting itself varies person by person. A general browser OCR tool should not pretend those tasks have the same reliability.
For a quick extraction, try Image OCR. For a scanned PDF, Make PDF Searchable keeps the visual scan while adding a text layer. In either case, the result is a draft until the important fields have been checked.