- Why the image decides OCR accuracy
- Seven reasons accuracy drops
- The order to change scan settings
- Narrowing capture for handwriting and small type
The short answer
When OCR results do not match what is on the page, the problem is almost always the image, not the software. Work through seven things — resolution, skew, shadows, handwriting, tables, small type and PDF compression — and most cases improve with a change of scan settings alone. Accuracy depends on the condition of each document, so the same settings will not give the same result on every kind of paper.
Why the image decides the outcome
OCR works by matching patterns of dark pixels to letter shapes. If there are too few pixels, if the shapes are distorted, or if there is little contrast between ink and background, the software sees a different character than you do.
Most "we changed the software and it still gets it wrong" cases start from the same image as before. Before you correct the output by hand, it is often faster to scan the page again.
Seven reasons accuracy drops
1. The resolution is too low
Many office copiers ship set to 200 dpi. Start by moving to 300 dpi. For documents full of 9-point type, or where you need every digit of a number, use 400 dpi. Higher is not automatically better: at 600 dpi the files get heavy and processing slows down.
2. The page went in crooked
A sheet that enters the feeder slightly at an angle produces lines that step sideways. Paper with staple holes or fold creases skews the most. Set the paper guides to the width of the stack, and feed about ten sheets at a time. If your scanner offers deskew or automatic orientation detection, turn it on.
3. Shadows and uneven lighting
Pressing a bound booklet onto the glass leaves a dark band down the spine. Photographing a page with a phone puts the shadow of your hand or body on the paper. Scan booklets one page at a time rather than as a spread, and when photographing, shoot from directly above with the light source in front of you rather than behind. Background removal settings reduce yellowing and show-through from the reverse side.
4. Handwriting is mixed in
Handwriting and printed type are not equally difficult. Rather than trying to capture handwritten addresses and free-text notes, limit capture to fields with a fixed format: dates, quantities, amounts. If you control the form, switching to one with a separate box per character changes the result on its own.
5. Table rules touch the text
When cells are narrow and characters touch the ruling lines, the line gets read as part of a character. Tables with many merged cells also lose the relationship between rows and columns. If you only need part of a table, crop that region and read it separately.
6. The type is small
Contract provisos, footnotes and marginal symbols are often two or three points smaller than the body text. Rescan just that area at a higher resolution, or take an enlarged copy first. When one page mixes several type sizes, trying to capture all of it in one pass takes longer than doing it in two.
7. The PDF is heavily compressed
"High compression PDF" or "smallest file size" on a copier saves the page with the character outlines already broken up. A setting chosen to make email attachments smaller then gets in the way of OCR. Keep compression light for the copy you will read, and produce a separate lightweight PDF for sharing.
Detail lost to compression cannot be restored afterwards by image processing. Decide the settings for your archive copy separately from the settings for sending.
The order to check things in
Changing everything at once makes it impossible to tell what helped. Change one setting at a time and re-test with the same page.
| Order | What you see | What to change |
|---|---|---|
| 1 | The file is unusually small | Reduce compression |
| 2 | Characters look blurred | Raise to 300–400 dpi |
| 3 | Lines run at an angle | Set the guides, enable deskew |
| 4 | Dark background, show-through | Background removal, grayscale (for archived originals, check the National Tax Agency's requirements) |
| 5 | Still wrong | Crop the region, or rescan |
How this looks in practice
Suppose an architectural practice of about ten people wants to file site inspection records that arrive on paper so they can be found later by property number. Say the original scans are 200 dpi high-compression PDFs of A3 forms captured as spreads, and almost none of the dates come through.
Changing to 300 dpi, lighter compression and one page at a time, and dropping the handwritten remarks column from capture, can narrow what goes unread to the handwriting. That makes filtering by date and property number easier.
Instead of capturing everything, capture only the fields you need in order to find the document later. That keeps the work down to a realistic amount of time.
Where to be careful
If captured values flow straight into a ledger, mistakes surface months later. Keep a step where a person checks the fields that matter downstream: amounts, dates and counterparty names.
If the stored documents have to serve as records for tax or regulatory purposes, separate requirements apply to resolution and colour depth, and greyscale may not be accepted for documents such as invoices. We do not cover those here; check the guidance published by the National Tax Agency.
Faded thermal receipts and worn carbon copies may not read at any setting. Rather than forcing them through, it is usually faster overall to decide that those go to manual entry.
In short
- The cause is usually the image, not the OCR engine.
- Change compression, resolution, skew and background in that order, one at a time.
- Narrowing what you capture makes results steadier on handwriting and small type.
- Keep a human check on the values that matter.
APOUL OCR Reader
APOUL PDF Split