Shot2Text

Why OCR Comes Out Messy — Six Causes and How to Fix Them

Almost every bad result is caused by the image, not the reader. Six problems, how to spot each one from the output, and the order to fix them in.

Short answer: almost every bad result is caused by the image, not the reader. Text recognition is a mechanical process — given clean edges it reads reliably, given smeared edges it guesses. Learning to spot the six problems below fixes more results than switching tools ever will.

What the reader is actually doing

It looks at the image, finds regions where dark pixels form letter-like shapes, segments them into lines and words, and matches each shape against a model of what characters look like.

That process fails in specific, predictable ways. Each section below is one of them, with the fix.

1. The text is too small

This is the most common cause by a wide margin.

The reader needs roughly 20 pixels of character height to be reliable and starts failing below about 12. Below that, a lowercase e is four or five pixels tall — there are not enough pixels to distinguish it from c or o, and the reader picks whichever the surrounding context suggests.

How to tell: the output has plausible words in the wrong places, or character substitutions that look like a small child's spelling.

The fix: zoom the page to 150–200 % before screenshotting, or move the camera closer so the text fills the frame. Do not upscale a small image afterwards — upscaling adds pixels but no information.

2. The contrast is too low

Text needs to be clearly darker than its background. When it is not, the reader cannot decide where letters end.

The usual culprits: light grey UI text, faded receipts, pencil, and photographs of paper taken indoors where white paper comes out mid-grey.

How to tell: whole words missing, or lines that fade out partway across.

The fix: capture the higher-contrast version — switch an app out of light-grey-on-white if it lets you, or move nearer a window. This site pre-processes photos with a contrast stretch that recovers most indoor-grey cases, but it cannot invent contrast that is not there.

3. The image is blurred or out of focus

A blurred edge spreads over several pixels, so adjacent characters bleed into each other. Two letters become one, or one becomes three.

How to tell: output that reads like rn where you know it said m, or spaces invented inside words.

The fix: retake the shot. Tap to focus on the text, hold still, and let the camera settle before pressing the shutter. For a screenshot there is nothing to fix — screenshots are never blurry, so if a screenshot reads badly the cause is something else on this list.

4. The page is tilted

The reader expects lines of text to run horizontally. A few degrees of tilt is enough to start losing whole rows near the edge, because the line drifts out of the horizontal band being scanned.

How to tell: the first few lines read well and later lines degrade, or the end of every line is garbled.

The fix: shoot straight down, square to the page. For an existing image, rotate it upright first. Phone cameras record orientation, so a sideways photo is straightened automatically — but a page photographed at an angle is tilted in the scene, and that cannot be fixed automatically.

5. There is shadow or glare

Shadows make one part of a page darker than another, so no single contrast setting suits the whole image. Glare is worse: a reflection is pure white, with no letters in it at all.

How to tell: a band of missing text matching where the shadow or reflection falls.

The fix: change your position relative to the light. Put the light behind or beside you, not in front. For glossy surfaces, move the light off-axis so the reflection lands outside the text. Never use flash on a document — it creates a hotspot in the middle and a dark falloff at the edges.

6. Letters do not stand apart

Joined-up handwriting, decorative script fonts, and tight letter spacing all remove the gaps the reader uses to find character boundaries.

How to tell: output that is a plausible-looking sequence of random characters.

The fix: there is not one, for cursive. Cropping to a single word sometimes helps because the reader has fewer alternatives to consider. For decorative fonts, try anyway — bold simple shapes often survive — and expect to correct the result.

Problems that are not the image

The layout has columns. The reader scans line by line across the full width, so two columns merge into one stream of text. No setting fixes this; crop one column at a time.

The document contains tables. The output is words in reading order, with no concept of cells. The values are all there, the structure is not.

The text is in a language you have not selected. This is the one failure that looks like a total breakdown rather than a partial one — the output is not garbled text in your language, it is convincing-looking nonsense in the wrong one, because the reader is matching every shape against an alphabet that does not contain it. Selecting the right language fixes it completely.

The output has extra punctuation. Outlines, drop shadows and watermarks around text get read as strokes. Crop tighter so they are smaller relative to the letters.

Reading the output critically

A useful habit: when a result looks wrong, work out which of the six causes explains it before trying anything.

  • Whole words missing, lines fading → contrast
  • Plausible words in wrong places, character swaps → size
  • First lines fine, later lines bad → tilt
  • A band of nothing → shadow or glare
  • Random character soup everywhere → wrong language, or cursive
  • Merged words, invented spaces → blur, or letter spacing

That diagnosis takes ten seconds and usually points at a fix that works on the first retry, instead of trying three tools that all fail for the same reason.

What this tool does automatically

Three fixes are already applied before the reader sees your image, so you do not have to think about them:

Contrast stretch on photos — finds the darkest and lightest parts of your actual image and spreads them across the full range. Fixes most indoor-grey cases.

Upscaling and sharpening on small JPGs — recovers much of what JPEG compression took away near letter edges.

Flattening transparency onto white — a transparent PNG would otherwise be drawn against black, making dark text invisible.

Everything else on this page is on you, at the moment you take the picture.

The order to try things

When a result is poor, fix them in this order — it is ordered by how often each one is the real cause:

  1. Crop closer to the text.
  2. Check the language selector.
  3. Zoom before capturing, if it was a screenshot.
  4. Rotate upright and reshoot, if it was a photo.
  5. Change the lighting, if it was a photo of paper.
  6. Find a cleaner copy of the image.

Changing tools comes after all six. Because every browser-based reader is doing the same work on the same pixels, a different tool will fail in the same way — and you will have spent ten minutes finding that out.

One case breaks the pattern and deserves its own answer: a table. Rows come out as flat lines with the columns collapsed, which looks like a recognition failure but is not — the characters are right, the structure is what went missing. Extracting a table from a screenshot covers the crop-per-column trick and the digit confusions that quietly wreck a column of numbers.