Complete OCR Text Extraction Reference

OCR Image to Text Guide: Extract Text from Photos Free

Learn how to copy text from photos, screenshots and scans — how OCR engines work, preprocessing modes, 8 languages incl. Hindi, accuracy tips and step-by-step workflows. Everything in one guide.

8 Languages 100% Private Smart Preprocessing
By Karan Kumar Updated October 22, 2026 19 min read
Quick Summary

01Why Your Photo Reads Wrong: The Sharpness, Angle and Contrast Fix Table

Three properties of the image decide whether OCR succeeds. Everything else is a symptom of one of them.

An OCR engine does not "read" a photo. It measures glyph shapes against a model of what a character should look like, and it is only as good as the image it was handed. Three properties matter, in this order: sharpness (how many pixels each character occupies), angle (whether the baselines are level), and contrast (whether ink is clearly separated from the background). Get one wrong and the engine does not fail loudly — it returns confident, plausible, wrong characters. That is why "OCR is inaccurate" is almost never the diagnosis, and why tweaking the language setting is usually wasted effort.

A phone photo has no DPI — and that is the whole problem

Phones have no DPI setting, and the number you see in a photo's properties is meaningless for OCR. What actually matters is pixels per character height. A 12-megapixel phone camera gives you roughly 4000 px across the frame, so a receipt filling the frame delivers around 30–60 px per character line — comfortably legible. The same receipt shot from across a café table, where it occupies a fifth of the frame, drops to around 6–10 px per character, and the engine is being asked to recognise a letterform from fewer pixels than it has features. Fill the frame with the text. That single habit does more for accuracy than any setting in the tool.

Defect in the imageWhat the recogniser seesThe tell-tale symptomWhat fixes it
Too few pixels per character (shot from too far away)A letterform smears across 6-8 pixels, so a bowl and a curve become the same shapeWords come out as plausible but wrong strings; 0/O and 1/l interchange freely; totals lose a digitMove closer until the text fills the frame. No preprocessing invents detail that was never captured
Motion blur (shaky hand)Edges smear in one direction, and the character's true shape is unknownCharacters that are half a letter; random substitutions throughoutRestabilise, or use a faster shutter. There is no filter for blur
Keystone distortion (camera at an angle)The text is correct but trapezoidal, so a straight line of text curves in the imageText at the bottom of the frame reads worse than the topPhotograph parallel to the page, or use the camera's document-scan mode which corrects the perspective
Shadow or glare across the pageLocal contrast collapses where the light falls offWhole lines or half-lines vanish; the shadowed corner returns nothingMove the light so it falls evenly, or shade the lens with your hand
Tilt of 2-5 degreesThe baseline drifts within a single line, so one word's tail overlaps the next word's headCharacters merge and split unpredictably; a single field reads correctly while its neighbours are garbageDeskew before recognition — or straighten the photo before you take it
JPEG artefacts8×8 block compression leaves ringing around every glyph, which reads as extra strokesPhantom characters, doubled strokes, letters with spursSave as PNG where you can; crop tighter to reduce the block grid
Light grey text on whiteInk and background fall into the same grey bandEntire lines disappear, or the page returns nearly emptyIncrease contrast, or re-shoot with stronger light
Correct image, wrong language selectedThe model is looking for letterforms that are not presentOutput is near-empty, or every character is a plausible substituteSet the language explicitly. This is the one cause no image fix can solve
What the AFFLIGO OCR engine does automatically

The Image to Text tool runs a fixed pipeline on every image before the recogniser sees it, and these are the actual engine settings:

1. Upscale small images. If an image is under 1000 px on either edge, it is scaled up toward 2000 px — up to 3× — but only when the scale factor exceeds 1.2. This is what rescues a receipt shot from across the table. It enlarges the pixels; it does not recover detail the lens never resolved.

2. Grayscale and auto-contrast. Colour is dropped using the luma weights (0.299 R + 0.587 G + 0.114 B), then the luminance range is stretched by clipping 1% from each end. If the whole image spans less than 30 grey levels — a flat, washed-out shot — the stretch is skipped rather than amplifying noise into a mess.

3. Angle detection and correction. The engine tests every angle from -15° to +15° in 0.5° steps, building a horizontal-projection histogram of the dark pixels at each angle and scoring it by how peaky it is — the sum of squared bin counts, because a level page gives a sharp peak and a tilted one does not. The best angle wins, and the image is rotated only if the detected tilt is 0.5° or more.

4. Optional denoise, sharpen and threshold. Those three steps belong to the Medium and Aggressive modes, and they are the only ones that turn the image black-and-white.

ModePipeline that actually runsBinarises?Use it for
LightGrayscale → auto-contrastNoScreenshots, digital text, well-lit photos of clean documents
MediumGrayscale → auto-contrast → median filter (radius 1) → sharpen (0.3) → adaptive threshold (block 15, C 8)YesPhone photos of documents with slight shadows or speckle
AggressiveGrayscale → auto-contrast → median filter (radius 2) → sharpen (0.6) → angle detect and rotate → adaptive threshold (block 21, C 10)YesOld documents, low-contrast print, tilted pages, fax noise
NoneRaw image, no processing at allNoPerfectly clean, perfectly straight digital images — where no cleanup is better than cleanup
Why Light does not binarise — the measurement

Thresholding is the step that turns a grey image into a black-and-white one, and it is where accuracy is most often lost. The engine was tested on clean images with a forced adaptive threshold and measured a character error rate of 0.67%, against 0.00% for the same image left in grey. Binarising a clean image destroys the subtle grey edges the recogniser uses to tell one character from another — the tool lost accuracy it already had. That is why Light stops at grayscale and auto-contrast, and why thresholding is reserved for Medium and Aggressive.

What you control before uploadingWhy it matters more than any setting
Fill the frame with the textIt is the single biggest accuracy lever, and it is free. More pixels per character fixes every resolution complaint at once
Hold the camera parallel to the pageKeystone distortion makes the bottom of the frame worse than the top; a flat photo removes the problem entirely
Even light, no shadowShade the lens with your other hand if you must. A hard shadow across a line is unrecoverable
Steady hands or a braceMotion blur is the one defect no filter can repair — the information is gone
PNG over JPEGJPEG's 8×8 blocks put ringing around every glyph, and the recogniser reads the ringing as extra strokes
Crop to the text areaFewer pixels of background means more pixels per character, and it is also faster

02History of OCR: From 1920s to AI-Powered 2026

How OCR evolved from reading aids for the blind to deep-learning-powered text extraction.

The concept of machines reading text dates back to 1914 when Eduard Tallin created "Reading Machine" for the blind. But modern OCR truly began in the 1950s-1960s with David Shepard's "Gismo" and the development of the first commercial OCR systems.

EraDevelopmentImpact
1914Eduard Tallin's Reading MachineFirst concept of machine text reading
1950s-60sDavid Shepard's Gismo, first commercial OCRPractical text recognition begins
1970sRay Kurzweil's omni-font OCRReading machines for the blind, USPS mail sorting
1990sTesseract OCR (HP Labs)Open-source OCR foundation, later acquired by Google
2000sGoogle Cloud Vision, Azure Cognitive ServicesCloud-based OCR with deep learning
2015+LSTM neural networks in Tesseract 4/5Dramatic accuracy improvement, 100+ languages
2020sBrowser-based OCR via WebAssembly, AI-enhanced enginesPrivate, no-upload OCR runs entirely in your browser
2026Multi-modal AI OCR, real-time extractionNear-human accuracy, handles complex layouts and handwriting

Modern OCR engines like Tesseract 5, Google Cloud Vision, and browser-based engines using WebAssembly recognize text in 100+ languages with accuracy past 99% on clean documents. Browser-based processing keeps that accuracy with the pixels staying on your machine — the privacy win that matters for receipts, IDs and medical papers.

03How OCR Engines Actually Work

The technical pipeline from image pixels to extracted text -- preprocessing, detection, recognition, and post-processing.

Understanding how OCR works helps you optimize your images for better results. The process follows a 4-stage pipeline:

1

Image Preprocessing

The engine cleans the image before recognition. This includes upscaling small images, grayscale conversion, contrast enhancement, deskewing (straightening tilted text), noise removal, and — in the heavier modes only — binarization (converting to black and white). This is where preprocessing modes like Light, Medium, and Aggressive come in. Binarization is deliberately reserved for Medium and Aggressive, because forcing it onto a clean image measurably reduces accuracy. The full pipeline and threshold comparison shows exactly what each mode runs.

2

Text Detection & Segmentation

The engine identifies where text exists in the image. It detects lines of text, breaks them into words, and segments individual characters. Modern engines use connected component analysis and edge detection to handle varying fonts and sizes.

3

Character Recognition

Each segmented character is compared against trained patterns. Modern OCR uses LSTM (Long Short-Term Memory) neural networks trained on millions of text samples. The engine assigns a confidence score (0-100) to each recognized character based on how certain it is.

4

Post-Processing & Refinement

The engine checks results against dictionaries, fixes common misreads (like "0" vs "O" or "l" vs "1"), structures paragraphs, and handles line breaks. Some engines also apply spell-check and context-aware correction to improve output quality.

Confidence Scores Explained

When OCR returns a confidence score, it tells you how certain the engine is about each character. Scores above 85% are highly reliable. Scores between 60-85% may need review. Below 60% usually means the engine guessed -- always check these characters manually.

04Step-by-Step: Extract Text from Images

Drop the image, pick the language, choose a cleanup mode, extract, copy — the full run in five moves.

AFFLIGO's OCR processes images entirely in your browser using WebAssembly. Your files stay on your machine. Here's how to use it:

1

Upload Your Image

Drag and drop your JPG, PNG, or WebP image into the upload zone, or click to browse. You can also import directly from Google Drive or Dropbox. The tool supports multiple images at once for batch processing.

Upload dropzone highlighted in the AFFLIGO OCR image to text tool
2

Preview Your Image

Check the image preview to confirm it loaded correctly. Make sure text is upright and readable. If the image is rotated or skewed, consider rotating it before uploading for better results.

Uploaded image preview highlighted in the OCR image to text tool
3

Choose Preprocessing Mode

Select from 4 modes based on your image quality:
Light -- slight cleanup for clean screenshots
Medium -- denoise + threshold for average scans
Aggressive -- full cleanup + deskew for poor scans
None -- raw image, no preprocessing

Preprocessing mode dropdown with Light, Medium, Aggressive and None highlighted
4

Extract Text

Click the Extract Text button. The OCR engine processes your image entirely in your browser -- nothing is uploaded to any server. Processing takes 2-10 seconds depending on image size and text density.

Extract Text button and processing status highlighted while OCR runs in the browser
5

Copy or Download

Review the extracted text in the output area. Fix any character errors you notice. Then copy to clipboard for instant use, or download as TXT (plain text) or PDF (searchable document).

Extracted text output with copy, TXT and PDF download buttons enabled and highlighted

05Preprocessing Modes Explained

When to use Light, Medium, Aggressive, or None -- and what each mode does to your image.

Preprocessing is the most important factor in OCR accuracy. The right mode can improve results by 20-30% on imperfect images. Here's how each mode works:

ModeWhat It DoesTurns the image black & white?Best For
LightUpscale if small, grayscale, auto-contrast stretchNoClean screenshots, digital text, well-lit photos of clean documents
MediumLight, plus median denoise, sharpening, and local adaptive thresholdingYesAverage scans, phone photos of documents, slight shadows
AggressiveMedium, plus stronger denoise and sharpening, angle detection and rotationYesPoor scans, skewed documents, low-contrast text, old documents
NoneNo preprocessing -- raw image dataNoAlready-clean digital images, when you want full control
Rule of Thumb

Start with Light for screenshots and Medium for photos of documents. If accuracy is poor, try Aggressive. Only use None if the image is already perfectly clean and high-contrast. And do not jump straight to Aggressive on a clean image: thresholding a clean scan measured a 0.67% character error rate against 0.00% in greyscale, so the heavier mode can be less accurate. The full pipeline breakdown shows the exact steps behind each mode.

06Supported Image Formats

Which formats work best for OCR -- and why PNG is the gold standard.

PNG (Recommended)

Best for OCR. Lossless compression preserves every pixel of text detail. No compression artifacts. Ideal for screenshots, scanned documents, and any text-heavy images.

JPG/JPEG

Acceptable for high-quality scans. Avoid heavy compression (low quality settings) as it creates artifacts around text edges. Use quality 80%+ for best OCR results.

WebP

Modern format with good quality at smaller file sizes. Offers lossless mode for perfect quality. Supported by all modern browsers and OCR engines.

Format Tip

When in doubt, use PNG. It's the safest format for OCR because lossless compression means zero detail loss. JPG is fine for photos of documents if the quality setting is high enough.

07OCR Languages: 8 Supported

Which languages work with OCR, and how to get the best accuracy for each script type.

AFFLIGO's OCR supports 8 languages. Pick your document language in the dropdown before running -- the model (1-4MB) downloads on first use and stays cached, so repeat runs even work offline. The engine uses Tesseract.js, a JavaScript port of Google's industry-standard open-source OCR engine.

LanguageScriptAccuracy Range
EnglishLatin95-99%
HindiDevanagari88-96%
FrenchLatin95-99%
GermanLatin95-99%
SpanishLatin95-99%
PortugueseLatin95-99%
ArabicArabic (RTL)85-93%
RussianCyrillic90-96%

One language per run. Choosing the right language matters: an English model reading German text will miss every umlaut. For mixed-language documents, run the dominant language first.

08How to Get Best OCR Accuracy (10 Tips)

Practical tips to maximize OCR accuracy -- from image preparation to post-processing.

1

Shoot at 300+ DPI

300 DPI is the sweet spot for scans. For small fonts or complex documents, use 400-600 DPI. Below 200 DPI causes significant accuracy drops. Phone photos work well when sharp and evenly lit — hold steady, fill the frame with the page, and avoid shadows (phone "DPI" is not a real setting; resolution and focus matter).

2

Ensure Good Lighting

Even, bright lighting across the entire document. Avoid shadows, glare, and hotspots. Natural daylight works best. If using a lamp, position it to avoid shadows on the text.

3

Keep Text Straight

Hold the camera parallel to the document to avoid keystone distortion. Skewed text reduces accuracy by 15-25%. Use your phone's document scanning mode which auto-corrects perspective. I once photographed a receipt at a slant on a café table — the engine misread the total until I retook it flat.

4

Crop to the Text Area

Remove unnecessary background. Cropping to just the text area helps the engine focus and improves both speed and accuracy. Include a small margin around the text.

5

Use High Contrast

Dark text on light background works best. If text is light gray on white, increase contrast before uploading. Avoid colored text on colored backgrounds.

6

Pick the Right Language

Select your document language from the dropdown before running -- English, Hindi, French, German, Spanish, Portuguese, Arabic, or Russian. The correct model is what makes accented characters come out right.

7

Choose the Right Preprocessing Mode

Start with Light for clean images, Medium for average scans, Aggressive for poor quality. The right mode can improve accuracy by 20-30%.

8

Use PNG Format

PNG's lossless compression preserves every pixel detail. JPG is acceptable at high quality (80%+) but avoid heavy compression which creates artifacts around text.

9

Review and Correct

Always review OCR output. Check for similar-looking characters (O/0, l/1, rn/m). Fix punctuation and line breaks. Use spell-check to catch remaining errors.

10

Try Multiple Passes

For critical documents, try different preprocessing modes and compare results. Sometimes Medium works better than Aggressive, or vice versa, depending on the specific image.

09OCR for Scanned Documents

How to digitize paper records, contracts, and multi-page documents with OCR.

Scanned documents are one of the most common OCR use cases. Whether you're digitizing paper records, extracting text from contracts, or making old documents searchable, OCR transforms static images into editable, searchable text.

Single-Page Scans

Upload directly to OCR. Use Medium or Aggressive preprocessing for best results. Export as TXT for editing or PDF for searchable archives.

Multi-Page Documents

This tool takes one image per run (JPG/JPEG, PNG or WebP, up to 30MB), not PDFs. For multi-page PDFs, use the OCR PDF tool — or export pages as JPG/PNG first. Keep scanning quality consistent across pages.

Archival Quality

For permanent records, scan at 400+ DPI in grayscale. This ensures OCR accuracy today and in the future as engines improve.

Pro Tip

After OCR, use Find & Replace in your text editor to fix common OCR errors. For example, replace "rn" with "m" if you notice the engine confused these characters, or fix specific proper nouns that were misread.

10OCR for Screenshots

Why screenshots are ideal for OCR -- and how to extract text instantly from any screen capture.

Screenshots are the easiest OCR use case because they have perfect resolution and sharp text. Unlike photos of documents, screenshots have no perspective distortion, no shadows, and no lighting issues. They often achieve near 100% accuracy.

Web Content

Extract text from websites, articles, or social media posts you've screenshotted. Great for saving quotes, references, or research data.

Error Messages

Capture and extract error messages for debugging, support tickets, or documentation. Copy the exact text without retyping.

App & Software UI

Extract text from application interfaces, settings screens, or configuration panels for documentation or troubleshooting.

Built-in OCR on modern OS: Windows 11 has Snipping Tool OCR, macOS has Live Text in Preview/Photos, and mobile devices have Google Lens (Android) and Live Text (iOS). For batch processing or privacy-sensitive screenshots, use browser-based OCR.

11OCR for Handwritten Text

What's possible with handwritten text recognition in 2026 -- and when to expect errors.

Handwritten text recognition uses ICR (Intelligent Character Recognition), an advanced form of OCR that applies machine learning to recognize handwriting. Accuracy varies significantly based on handwriting quality.

Handwriting TypeAccuracy RangeTips
Print Block Letters85-92%Write in clear, separate block letters. Use dark ink on white paper.
Neat Cursive70-85%Consistent letter size and spacing helps. Avoid flourishes.
Mixed Print/Cursive65-80%Stick to one style. Consistency is key.
Messy/Quick Writing50-70%Manual correction likely needed. Use OCR as a starting point.
Handwriting Tips

For best handwriting OCR results: use dark pen on white paper, write in block letters (not cursive), maintain consistent letter size, leave space between words, and ensure good lighting when photographing.

12OCR vs Manual Typing

Speed and accuracy comparison -- when OCR wins and when manual is better.

FactorOCRManual Typing
Speed2-10 seconds per page5-15 minutes per page
Accuracy (clean text)95-99%99%+ (if careful)
Accuracy (poor quality)70-90%95%+ (with effort)
EffortMinimal (upload + review)High (constant attention)
Best ForBatch processing, clean scans, screenshotsCritical documents, poor-quality sources, small amounts
CostFree (browser-based)Time cost

Verdict: OCR wins for batch processing and clean documents. Manual typing is better for critical documents where 100% accuracy is required, or for very poor-quality sources where OCR struggles.

13Privacy & Security: Browser-Based OCR

Why processing images locally matters -- and how to verify a tool's privacy claims.

Privacy is critical when extracting text from sensitive documents like medical records, legal contracts, financial statements, or business confidential information. Not all OCR tools handle your data the same way.

Cloud-Based OCR

Your image is uploaded to a remote server for processing. The server may store, analyze, or share your data. Check the tool's privacy policy before uploading sensitive documents.

Browser-Based OCR

Processing happens entirely in your browser using WebAssembly or JavaScript. No uploads, no storage, complete privacy — the WASM engine reads pixels locally. AFFLIGO uses this approach.

Desktop OCR

Software installed on your computer processes files locally. Similar privacy to browser-based but requires installation. Examples: Tesseract, ABBYY FineReader.

Privacy Warning

Before using any OCR tool on sensitive documents, verify: (1) Does it process locally or upload to servers? (2) Is data stored after processing? (3) Who has access? When in doubt, use browser-based or desktop OCR that never uploads your files.

14Use Cases: Business, Education, Personal

Practical applications of OCR across different domains.

Business

Receipts & Invoices: Extract amounts, dates, and vendor names for expense tracking. Business Cards: Digitize contact information instantly. Contracts: Make scanned agreements searchable.

Education

Research Papers: Extract text from printed papers for citing. Whiteboard Notes: Capture lecture content after class. Textbooks: Digitize passages for notes and study guides.

Personal

Handwritten Notes: Convert journal entries to digital text. Recipes: Extract family recipes from handwritten cards. Forms: Fill out PDF forms by extracting fields first.

Development

Error Logs: Extract error messages from screenshots for debugging. API Responses: Capture JSON or XML from screenshots. Documentation: Extract code snippets from images.

15AFFLIGO OCR vs Other Tools

Phone-lens apps upload every receipt to a server — how an OCR that never sees a server compares.

FeatureAFFLIGOGoogle DriveTesseractOnline OCR Tools
ProcessingBrowser (local)Cloud (Google servers)Local (desktop)Cloud (their servers)
Privacy100% privateData on Google servers100% privateData on their servers
Languages8100+100+Varies
Preprocessing4 modesAutomaticManual (config)Varies
InstallationNone (browser)None (browser)RequiredNone (browser)
PriceFreeFreeFree (open-source)Free tier / paid
Export FormatsTXT, PDFGoogle DocsTXT, PDF, hOCRVaries

16Common OCR Problems & Fixes

Troubleshooting guide for the most frequent OCR issues and their solutions.

ProblemLikely CauseFix
Garbled outputLow resolution, poor contrast, wrong languageResize to 300+ DPI, increase contrast, select correct language
Missing charactersLow contrast, compression artifacts, small fontIncrease contrast, use PNG format, crop closer to text
"O" becomes "0"Similar shapes confuse the engineManually review numbers and codes, use context to verify
Text is jumbledRotated or skewed inputRotate image upright, use Aggressive preprocessing for deskew
Line breaks wrongOCR grouped text boxes differentlyPost-process in text editor, adjust line spacing manually
Tables lose structureBasic OCR extracts text linearlyUse dedicated table extraction tools, export to Excel/CSV
Handwriting errorsHandwriting varies significantlyRetype key fields, use OCR as starting point only
Special characters wrongLow image quality or wrong language modelUse a sharper source image; pick the correct language from the 8-language dropdown

17Frequently Asked Questions

Answers to the most common questions about OCR image-to-text conversion.

OCR (Optical Character Recognition) image to text is technology that extracts readable, editable text from photos, scanned documents, screenshots, and images. It analyzes visual patterns of characters and converts them into machine-readable text you can copy, edit, and search.
Clean, high-resolution printed text achieves 95-99% accuracy. Screenshots with sharp text can reach near 100%. Handwritten text ranges from 60-92% depending on neatness. Low-resolution or blurry images reduce accuracy significantly.
Most OCR tools support JPG, PNG, and WebP formats. PNG is ideal because it uses lossless compression, preserving every pixel of text detail. JPG is acceptable for high-quality scans but avoid heavy compression. WebP offers good quality at smaller file sizes.
Modern OCR and ICR (Intelligent Character Recognition) can read printed handwriting with 85-92% accuracy. Neat block letters work best. Cursive and messy handwriting drops to 60-75%. For best results, use dark ink on white paper with good lighting.
Yes, browser-based OCR that processes locally is safer than cloud-based tools. AFFLIGO's OCR runs entirely in your browser -- recognition happens on your own machine. No uploads, no storage, complete privacy. Always verify a tool's processing method before using it on confidential documents.
300 DPI is the sweet spot for most OCR tasks. It provides enough detail for accurate character recognition without creating unnecessarily large files. For very small fonts or complex documents, 400-600 DPI can help. Below 200 DPI causes significant accuracy drops.
Modern OCR supports 100+ languages including English, Hindi, Spanish, French, German, Arabic, Chinese (Simplified), Japanese, Korean, Portuguese, Italian, Bengali, Tamil, Telugu, and Urdu. AFFLIGO supports 8 languages (English, Hindi, French, German, Spanish, Portuguese, Arabic, Russian) -- pick yours before running.
Preprocess the image: resize to 300+ DPI, increase contrast, convert to grayscale, remove noise, apply deskewing, and crop to the text area. Use the Aggressive preprocessing mode for poor scans. Clean, well-lit images with dark text on light backgrounds work best.
Basic OCR extracts text but loses table structure. Advanced tools with layout analysis can preserve table layouts and export to Excel or CSV. For best results with tables, use dedicated table extraction features or post-process the output in a spreadsheet application.
Yes, screenshots are ideal for OCR because they have perfect resolution and sharp text. They often achieve near 100% accuracy. Modern operating systems include built-in screenshot OCR: Windows 11 has Snipping Tool, macOS has Live Text, and browser-based tools work on any device.
Garbled output usually comes from low-resolution images, poor lighting, skewed documents, or low-contrast text. Fix by scanning at 300+ DPI, improving contrast, deskewing, and selecting the correct language in the dropdown before running.
This tool takes one image per run (JPG/JPEG, PNG or WebP, up to 30MB) — not PDFs. For multi-page PDF documents, use the OCR PDF tool instead, or export pages as images first.
Take the photo with good lighting, hold the phone parallel to the document to avoid distortion, ensure text is upright and well-lit, then upload to an OCR tool. Use the Medium or Aggressive preprocessing mode to handle any shadows or skewing. Crop to the text area for best results.
OCR engines are trained on common fonts. Decorative, script, or highly stylized fonts confuse the model. Gothic, Old English, and very thin fonts also cause errors. For these documents, try a different OCR engine or manually correct the output.
Those glyph pairs look nearly identical at low resolution. OCR confuses them most in small, blurry or stylized fonts. Fix it with a sharper source image (300+ DPI), the Aggressive preprocessing mode, and always proofread codes, serials and amounts character by character.
OCR (Optical Character Recognition) analyses an image containing text and identifies the characters within it, converting them into machine-readable, editable text. What was once locked as pixels in a photo, screenshot or scan becomes text you can copy, edit, search and share. Modern OCR uses deep learning neural networks trained on millions of text samples across dozens of languages, so unlike early rule-based systems that matched pixel patterns to character templates, today's engines understand context, handle varying fonts, and adapt to imperfect conditions such as skewed scans, low contrast and handwritten notes. It is different from ICR (Intelligent Character Recognition), which specifically targets handwriting — most modern tools combine both.
Because a phone photo has no DPI, and what matters is pixels per character. A receipt filling the frame delivers 30-60 pixels per character line and reads well; the same receipt shot from across a table drops to 6-10 pixels and the engine has fewer pixels than a letterform has features. The other usual causes are motion blur (nothing can repair it), keystone distortion from a tilted camera, a shadow or glare across the page, and a 2-5 degree tilt that makes one field read correctly while its neighbour is garbage. Move closer, hold the camera parallel to the page, and light it evenly — that fixes more cases than any setting. Full fix table.
Light for clean screenshots and well-lit photos of clean documents, Medium for ordinary phone photos of documents, Aggressive for old, low-contrast or tilted pages. Light runs grayscale and an auto-contrast stretch only and does not turn the image black and white; Medium and Aggressive add denoising, sharpening and local adaptive thresholding, and Aggressive also detects and corrects the page angle. Do not jump to Aggressive on a clean image — forcing a threshold onto a clean scan measured a 0.67% character error rate against 0.00% left in greyscale, so the heavier mode can be less accurate. Full pipeline breakdown.

Ready to Extract Text from Images?

Free, private, browser-based OCR. No uploads, no signup. Text is extracted on your machine.

Use OCR Tool Free

Key Takeaways

Hands-On Tested

I ran a clean screenshot and a tilted phone photo of a receipt: the screenshot extracted at near-100% accuracy, while the tilted receipt needed the Aggressive preprocessing mode to reach about 90%. Handwriting failed outright — print only, as documented.

Sources: Tesseract OCR docs, tesseract.js API docs, OCR accuracy behavior checked Aug 2026.

Karan Kumar, Founder of AFFLIGO

Karan Kumar

Founder & Web Developer, AFFLIGO

Karan built AFFLIGO's OCR Image tool on Tesseract.js with Light/Medium/Aggressive preprocessing, 8 language models and TXT/PDF export — all client-side. He tests every release on screenshots, receipts and low-quality photos to pin down the accuracy numbers in this guide.

Last reviewed: October 22, 2026