Experiment

Why Does Photo OCR Misread Text? A Small Size-and-Blur Test

We downsized and blurred the same printed English and Korean samples, then compared OCR output. See when to retake a photo and when to correct a few characters.

When extracted text is wrong, it is tempting to blame the recognition model alone. Yet the same engine can return different text depending on whether the letter strokes survive resizing or blur.

We used Toolit's Korean-and-English OCR model on two short printed test images, changing image size and blur. This is a controlled illustration, not a prediction of accuracy across phone cameras and documents.

What this test showed

Smaller was not always wrong

The English sample remained correct at half and quarter of its original dimensions.

Very small text changed

At 90×23 pixels TOOLIT became TOOUIT; at 95×24 the Korean 툴릿 became 플릿.

Strong blur removed the result

The heavily blurred English sample returned no text. A clear source matters.

We resized or blurred identical printed words

  1. 1

    We used two non-personal PNG fixtures from the repository: a 720×180 English image reading ‘HELLO TOOLIT 123’ and a 760×190 Korean image reading ‘안녕하세요 툴릿 123’.

  2. 2

    We recognized the originals and transformed copies with the same Tesseract Korean+English language data used by the tool. English copies were 360×90, 180×45, and 90×23; Korean had a 95×24 copy. We also made mildly and heavily blurred English copies.

  3. 3

    We compared returned text, not a statistical accuracy score. The reproducible script is scripts/measure-ocr-and-alpha.mjs. Each controlled variant was run once; this is not an average of real photographs.

What the engine actually read

Moderate downsizing and mild blur did not change this sample's English output. More severe downsizing altered English and Korean words; heavy blur yielded no English text. This does not mean every image fails below a particular dimension. The useful detail depends on the original lettering and image conditions.

InputReturned textCompared with source
English original 720×180HELLO TOOLIT 123Same
English 360×90 / 180×45HELLO TOOLIT 123Same
English 90×23HELLO TOOUIT 123TOOLIT misread
English, heavy blurEmpty resultUnreadable
Korean original 760×190안녕하세요 툴릿 123Same
Korean 95×24안녕하세요 플릿 123"Name misread; extra mark

Why enlarging a bad source may not help

Small lettering compresses the gaps and short strokes that distinguish characters into just a few pixels. Blur merges their edges with neighboring pixels. OCR then guesses from the remaining shapes and language clues, so similar letters or unexpected punctuation can appear.

Enlarging an already tiny copy cannot restore missing strokes. Tesseract's own quality guide discusses resolution, skew, noise, and image preparation. Better input helps, but it does not guarantee every name or digit will be right.

Retake or correct?

If most text is sound and only a few letters are wrong, correcting the editable result against the image is usually fastest. If entire lines are missing or repeatedly nonsensical, start with a new photo or screenshot. Avoid glare, keep the sheet flat, include the full text, and select the relevant language.

Check every high-impact figure such as an amount, date, or account number even if the rest looks accurate. Plain text also needs manual review when a table's cell relationships matter.

Limits of this observation

These were two short printed lines, not handwriting, business cards, rotated receipts, tables, or text on a busy image. We did not test every camera or image-compression method. The transformation sizes are not universal pass/fail thresholds.

For an actual document, placing the source beside the extracted result and correcting it remains the last quality-control step.

Official documentation used for this analysis

Start with a readable source; verify the result

Severe resizing or blur changed or erased text in this controlled test, whereas a milder reduction did not. There is no useful one-size-fits-all pixel cutoff; inspect the letter shapes and the actual output.

Retake a blurred source where possible. Then compare important names and numbers against the image. OCR saves typing time, not the need for review.