OCR Accuracy: Fix the Input Before the Engine

When OCR results are poor, the first instinct is to blame the OCR engine and look for a better one. Often the real problem is upstream: a phone photo taken at an angle, a faxed form scanned at low resolution, a grey background that confuses the conversion to black and white. OCR accuracy depends as much on the images you feed in and how you measure the results as on the engine itself. This guide explains how to measure accuracy in a way that reflects business impact, which image problems cause most errors, how to fix them before recognition, and how to handle the errors that remain.

Measure OCR accuracy at the field level

Accuracy can be measured in several ways, and the choice matters. Character-level accuracy, the share of characters recognised correctly, sounds precise but hides what matters. A page can be almost perfectly recognised while the one wrong character is in the invoice total.

For business documents, measure at the field level:

• Field accuracy: the share of key fields, such as invoice number, date and total, that are exactly right
• Document accuracy: the share of documents where every key field is right
• Error by field: which fields fail most often

Build a test set of real documents, a few hundred if you can, that represents your actual mix of sources, layouts and quality. Record the correct values by hand once. Every change to preprocessing, settings or engine is then measured against the same set, so you know whether it helped.

Why image quality drives OCR accuracy

Recognition engines are trained on reasonably clean, straight, well-lit text. Real documents depart from that in predictable ways. The open-source Tesseract project's guide to improving output quality is a useful reference, because the problems it describes apply to most engines:

• Resolution: the guide says Tesseract works best on images with a DPI of at least 300, so low-resolution images may benefit from resizing
• Binarisation: converting to black and white is done internally, but the guide notes the result can be poor when the page background is unevenly dark
• Skew: the guide says line segmentation quality drops significantly when a page is too skewed, which severely affects recognition
• Borders: text cropped with no margin can cause problems, and the guide suggests adding a small border
• Noise: specks, stamps and lines across text are read as characters or break words

Fix images before recognition

Preprocessing is the set of steps that correct these problems before the engine sees the image. Common steps:

1. Convert to a consistent format and colour mode
2. Resize low-resolution images towards the recommended resolution
3. Correct rotation and skew so text lines are horizontal
4. Remove noise, such as speckles and scanning artefacts
5. Binarise with a method suited to uneven backgrounds; the Tesseract guide notes that version 5 added adaptive methods for this
6. Crop to the document edges and add a small margin

Apply preprocessing selectively. Clean digital PDFs often contain a text layer already and need no OCR at all; extract their text directly. Photos from phones usually need the most correction.

Fix the capture where you can

The cheapest improvement is often at the source. If staff scan documents, set scanners to a sufficient resolution and grayscale rather than low-quality black and white. If customers upload photos, guide them in the upload screen: flat surface, good light, whole page in frame. If suppliers send faxes or poor scans, ask for digital PDFs. A small change in capture can remove a whole class of errors.

Use layout and context

Even with perfect characters, extraction can fail if the engine misreads the layout, such as reading two columns as one line, or mixing table cells. Settings that tell the engine what layout to expect help. Tesseract, for example, expects a full page of text by default, and the same guide suggests choosing a different page segmentation mode when recognising only a small region.

Context also corrects errors after recognition. An invoice date must be a valid date; a total must be a number; a supplier name should match a known supplier. Validation rules turn many recognition errors into flagged items rather than wrong data. Our guide to Python invoice parsing shows how parsing and checks work on extracted text.

Worked example: diagnosing low OCR accuracy

The table below shows a hypothetical diagnosis for a team processing delivery notes from many sources. Error patterns and fixes are illustrative; measure your own.

Symptom in test setLikely causeFixHow to confirmPhone photos have many wrong charactersSkew and perspectiveDeskew and perspective correctionField accuracy on photo subsetFaxed notes read as gibberishLow resolutionResize before recognition; request PDFs from senderField accuracy on fax subsetGrey forms lose faint textPoor binarisation on uneven backgroundAdaptive binarisationCompare before and after on grey formsTotals merged with adjacent textLayout misreadAdjust layout settings or crop regionsTotal field accuracyStamps over text break wordsNoiseNoise removal; review flag when stamp detectedError rate on stamped documentsClean PDFs still have errorsOCR used on files with a text layerExtract the text layer directlyAccuracy on digital PDFs

Each fix is tested on the relevant subset, so improvements in one area do not hide regressions in another.

Handle the errors that remain

No preprocessing removes every error. Plan for the remainder:

• Use confidence scores where the engine provides them, and route low-confidence fields to review
• Apply validation rules on every key field
• Record corrections, so you can see which sources or layouts cause most errors
• Consider a different engine or extraction service only after measuring that image quality is not the cause

Our guide to intelligent document processing explains how recognition, validation and review fit together into one pipeline.

Common mistakes when chasing OCR accuracy

• Measuring character accuracy instead of the fields that matter
• Testing on a handful of clean samples instead of the real mix
• Switching engines before fixing image quality
• Running OCR on digital PDFs that already contain text
• Applying heavy preprocessing to every image, degrading clean ones
• No validation, so recognition errors become data errors

Frequently Asked Questions

How do you measure OCR accuracy?

For business documents, measure field accuracy on a representative test set: the share of key fields, such as dates and totals, that are exactly right, and which fields fail most often.

What resolution is best for OCR?

The Tesseract project's guide says Tesseract works best on images of at least 300 DPI. Check the guidance for the engine you use.

Why do photos of documents give poor OCR results?

Photos are often skewed, unevenly lit and low in contrast. Deskewing, perspective correction and adaptive binarisation usually help.

Will a better OCR engine fix accuracy problems?

Sometimes, but often the cause is image quality, layout or missing validation. Measure the effect of preprocessing and capture changes before switching engines.

Improve extraction where it counts

We diagnose OCR accuracy problems with real test sets and fix the pipeline from capture to validation. See our AI automation solutions, or talk to us about your documents.

About the Author

Manav Sharma is the founder of Insyrge and a CRM solutions architect. He works across Zoho, Salesforce, HubSpot, Odoo and more, designing the CRM setups, integrations and automations that keep sales and operations teams working from the same data.