Why Text Extraction From Photos Goes Wrong, and How to Fix It
How text size, blur, tilt, shadows, JPG quality and the wrong language setting change OCR accuracy, measured on the same engine Hanji uses.
You photograph a receipt and the total comes back garbled, while a screenshot of the same receipt reads perfectly. A form scanned straight is fine, yet the same form photographed at a slight angle comes out as nonsense. Reading text from a picture means guessing every character from its pixels, so a small drop in input quality can push the result over a cliff instead of lowering it gently. If you know where the cliffs are, you know how to take the photo and which settings to choose.
This guide is built on measurements, not general advice. We fed test images to the same engine that powers Hanji’s Image to Text and recorded how much each problem costs.
The short version
- Each character needs to be about 12 pixels tall. Japanese fell to 85% at 10 pixels and 38% at 8. English was still at 98% at 8 pixels.
- Blur is a cliff, not a slope. At 24-pixel text, 2 pixels of blur still read at about 91%. At 3 pixels, Korean and Japanese dropped to 11–12%.
- Keep the tilt under 3 degrees. At 7 degrees Japanese read 40%, Korean 84% and English 94%. At 15 degrees Korean was down to 25%.
- Leave “Even out lighting and contrast first” on when there are shadows. On a page with one dark half, it scored 95–100% with the option on and 39–57% with it off.
- JPG quality only matters when the text is small. With large text, quality 8 still read mostly fine. With small text, quality 20 took Japanese down to 9%.
- The wrong language gives you nothing. Reading Japanese as Korean scored 0%. Vertical Japanese needs “Japanese, vertical” (0–6% with the horizontal setting, 97–100% with the vertical one).
How we measured
On 23 October 2026, in Chrome on Windows, we put test images into Hanji’s Image to Text page and read back the text it produced. The engine is tesseract.js 7.0.0 with the tessdata_fast language data this site serves for Korean, Japanese, English and vertical Japanese. Settings were left at their defaults, so for Korean and Japanese “Also read English words and numbers” was on.
- Test images. Dark grey text on a white background, drawn on a browser canvas, so these are synthetic images. Fonts were Malgun Gothic (Korean), Yu Gothic (Japanese) and Arial (English), with three or four lines per image. At 24-pixel text the images were 628×181 (Korean), 605×181 (Japanese) and 467×222 (English).
- Sentences. Two different sets of sentences per language, averaged. The English set is receipt lines with an order number, dates, prices and totals. The Korean set has dates, order numbers, amounts and an email address. The Japanese set is notices.
- Accuracy. The share of the reference characters that came out right: 1 minus the edit distance (characters deleted, inserted or changed) divided by the number of reference characters. Spaces are ignored, and full-width and half-width forms count as the same.
- Text size. The font size in pixels. Korean and Japanese characters fill most of their square, so this is close to the character height. English lowercase letters are shorter. Before reading, the tool enlarges any image whose long side is under 1000 pixels by 2× and under 1800 pixels by 1.5×, so the numbers below include that.
These numbers have limits. They come from synthetic images, so real phone photos with perspective distortion, crumpled paper, glare and bad focus are not covered. Handwriting was not measured either. The values are also averages of two sentence sets and they differ more than you might expect near the edge: at 10 pixels, Korean scored 95% on one set and 75% on the other. Trust the big drops in the tables, not differences of a few points.
Results
Text size
| Text size | Korean | Japanese | English |
|---|---|---|---|
| 8 px | 88% | 38% | 98% |
| 10 px | 85% | 85% | 100% |
| 12 px | 95% | 99% | 100% |
| 16 px | 96% | 99% | 100% |
| 24 px | 95% | 100% | 100% |
| 48 px | 93% | 100% | 100% |
Japanese changes most steeply between 10 and 12 pixels. English held up at small sizes, because its letter shapes are simpler. Korean stayed around 95% even at large sizes, and that was not a size problem: a few words were misread every time.
Blur (24-pixel text, Gaussian blur, standard deviation)
| Blur | Korean | Japanese | English |
|---|---|---|---|
| None | 95% | 100% | 100% |
| 1 px | 97% | 99% | 100% |
| 2 px | 91% | 91% | 94% |
| 3 px | 11% | 12% | 48% |
| 4 px | 6% | 2% | 5% |
At 2 pixels the text is still readable and at 3 it is gone. There is no gentle decline in between. These are values for 24-pixel text only, and we did not measure where the edge sits at other sizes.
Tilt (24-pixel text)
| Tilt | Korean | Japanese | English |
|---|---|---|---|
| 0° | 95% | 100% | 100% |
| 3° | 99% | 95% | 99% |
| 7° | 84% | 40% | 94% |
| 15° | 25% | 5% | 84% |
| 25° | 15% | 2% | 35% |
English lines are long and easy to follow when they lean. Korean and Japanese were far more sensitive, and Japanese was already below half at 7 degrees.
One half in shadow (24-pixel text)
From the middle of the image towards the right, the paper was multiplied down to about 37% of white (grey 95).
| Lighting fix | Korean | Japanese | English |
|---|---|---|---|
| On (default) | 95% | 100% | 100% |
| Off | 56% | 39% | 57% |
JPG quality
Images saved as JPG at 24-pixel and 12-pixel text, with quality 90, 50, 20 and 8 (the 0–100 scale Chrome uses when saving from a canvas).
24-pixel text
| Quality | Korean | Japanese | English |
|---|---|---|---|
| 90 | 95% | 100% | 100% |
| 50 | 97% | 99% | 100% |
| 20 | 98% | 100% | 100% |
| 8 | 95% | 91% | 99% |
12-pixel text
| Quality | Korean | Japanese | English |
|---|---|---|---|
| 90 | 97% | 96% | 100% |
| 50 | 95% | 94% | 99% |
| 20 | 84% | 9% | 97% |
| 8 | 46% | 5% | 76% |
What did not cause trouble
- Faint text. Light grey text (#999999, #bbbbbb, #cccccc) on white scored 92–97% in Korean, 98–100% in Japanese and 100% in English.
- Light text on a dark background. Pale grey on near-black scored 99% (Korean), 98–99% (Japanese) and 100% (English), with the lighting fix on or off.
- Handwriting-style fonts. An English receipt set in Ink Free and Segoe Script read 98% and 95%. These are fonts, not real handwriting. We did not measure handwriting, so do not count on it.
Fixing each cause
Text that is too small
Move closer before you take the photo, so the text fills the frame, and check that your phone camera is not set to a low resolution. For scans, work from the resolution: 10-point type (about 3.5 mm) is about 21 pixels tall at 150 dpi and about 10 pixels at 72 dpi. The fine print on a receipt is closer to 6 points, which is only about 12 pixels at 150 dpi, right at the edge, so scan it at 300 dpi. These two figures are arithmetic against the 12-pixel mark above, not measurements. The pixels-and-dpi relationship is explained in Pixels, cm and DPI: how to convert, with tables.
Enlarging a small picture yourself and reading it again rarely helps. The tool already enlarges small images up to 2× before it reads them, and enlarging does not create detail that was never there.
Blur
Blur cannot be undone, and as the table shows, the drop is sudden. Tap the text to focus and keep the phone still as you press the shutter. In dim light the phone leaves the shutter open longer and your hand shake gets recorded, so moving to a brighter spot is the easiest fix. If a photo is already blurred, retaking it is quicker than trying to rescue it.
Tilt
Tilt is fixable. Open Crop Image, use “Straighten” until the lines are level, trim the edges, and then extract the text. When you shoot, hold the phone flat above the page so that it is parallel to the paper, which reduces both tilt and perspective distortion. Our tilt test was a plain rotation, so we did not measure the keystone effect you get from shooting at an angle, where the top lines look small and the bottom lines large.
Shadows and uneven light
“Even out lighting and contrast first” is on by default. In the test above it lifted a half-shadowed page from 39–57% to 95–100%. If a clean screenshot reads worse with it on, turn it off. We did not run a separate comparison of on and off for images without shadows. To avoid the shadow in the first place, keep your own body and phone from blocking the light and shoot by a window.
Photos saved as JPG several times
A photo that has been through a messaging app has usually been recompressed. With large text it makes little difference, but small text falls apart suddenly, as the table shows. For documents with fine print, use the original instead of the copy you received in chat. Also avoid lowering the quality with Compress Image before you extract the text. Shrinking the file is fine once the text has been extracted.
Choosing the language, and vertical text
The tool cannot tell which language the picture is in. It searches only among the characters of the language you pick, so the wrong choice gives you meaningless output. The table shows the same image read with different language settings. Each row uses one sentence set (24-pixel text).
| Image | Language chosen | Accuracy |
|---|---|---|
| Korean receipt or notice | Korean + English | 92% |
| Korean receipt or notice | Korean only | 75% |
| Korean receipt or notice | English only | 47% |
| Korean receipt or notice | Japanese + English | 55% |
| Korean and English mixed on one line | Korean + English | 98% |
| Korean and English mixed on one line | Korean only | 69% |
| Korean and English mixed on one line | English only | 58% |
| Japanese notice | Japanese + English | 100% |
| Japanese notice | Korean + English | 0% |
| Japanese notice | English only | 0% |
| English receipt | English | 100% |
| English receipt | Korean only | 38% |
| Vertical Japanese | Japanese + English | 0–6% |
| Vertical Japanese | Japanese, vertical | 97–100% |
- For Korean documents with numbers or English words, keep “Also read English words and numbers” on. With Korean only, a document with an order number and an email address scored 69–75%. Reading English too brought it to 92–98%. With English only, the Korean could not be read at all, and the result was 47–58%.
- For English forms and receipts, choose English. The English receipt scored 100% with settings that also read English (Korean + English, Japanese + English), but only 38% with Korean and no English.
- Choose the language of the document. Reading a Japanese notice as Korean did not get a single character right.
- For vertical Japanese, choose “Japanese, vertical”. With the horizontal setting the engine takes the lines the wrong way and reads almost nothing (0–6%). The reverse does not work either: horizontal text read as vertical scored 6%. Keep pages that mix both directions apart and read them separately. When you choose vertical, the English option disappears, so check any Latin letters and digits in vertical text by eye.
Scanned PDFs: use PDF to Text
Every page of a scanned PDF is a picture. If a PDF looks like text but you cannot select it, that is what you have. You do not need to turn pages into images one by one. Open the file in PDF to Text: pages that already contain text are copied out, and pages that are only a picture are read. The results above carry over to that route. When you scan, use 150 dpi or more so that the text ends up at least 12 pixels tall. Scanning tips and ways to keep the file small are in Scan Documents to PDF with Your Phone.
Privacy
Receipts, forms and ID photos contain names, addresses and account numbers. Image to Text downloads the engine and language data from hanji.page and reads the picture inside this browser tab, so the image is not uploaded to a server. All the measurements in this guide also ran on one PC and nothing left it. How to check any tool yourself is covered in Do Online File Converters Keep Your Files?.
Related tools
- Image to Text: pull the text out of a photo or screenshot, copy it, or save it as a .txt file.
- PDF to Text: extract text from a PDF, reading scanned pages when there is no text layer.
- Crop Image: straighten a tilted photo and trim the edges.
Guides
- Photo too large to upload? How file size limits work
- HEIC photos from iPhone: why forms reject them and fixes
- JPG, PNG, WebP, AVIF and HEIC compared
- Pixels, cm and DPI: how to convert, with tables
- How to Make a PDF Smaller (and What You Give Up)
- Scan Documents to PDF with Your Phone
- ID and passport photo sizes in mm and pixels
- How to Take a Passport Photo at Home That Gets Accepted
- Photo file rules for online applications: pixels, KB and age
- Do Online File Converters Keep Your Files?
- Print Passport Photos on 4×6: CVS, Walgreens, Walmart
- Character and Word Counts: Why Counters Disagree
- QR Codes That Keep Working: Static, Dynamic and Print
- Email Attachment Size Limits: Gmail, Outlook, Yahoo, iCloud
- Shrinking Images to a File Size: What We Measured