Learn how to extract text from scanned PDFs and images using OCR. Step-by-step guide covering online tools, accuracy tips, language support, and best practices for document digitization.
Free, in your browser — no signup, files auto-delete in 2 hours.
Optical Character Recognition (OCR) is the technology that converts images of text into actual, editable, searchable text data. When you scan a paper document or receive a PDF that was created from a scanned image, the text you see on screen is not really "text" at all -- it is a picture of text. You cannot select it, copy it, search within it, or edit it. OCR bridges that gap by analyzing the visual patterns in the image and recognizing individual characters, words, and paragraphs.
The practical applications are enormous. Businesses digitize paper archives to make decades of records searchable. Law firms process thousands of scanned contracts for e-discovery. Researchers extract data from historical documents. Individuals convert printed recipes, handwritten notes, or photographed whiteboards into editable text. Healthcare organizations digitize patient records. Government agencies make public documents accessible.
This guide walks you through the process of OCR -- from preparing your source documents to extracting accurate text, with specific attention to the tools, settings, and techniques that produce the best results.
Modern OCR systems process documents through several stages:
Preprocessing. The image is cleaned up: deskewed (straightened), denoised (speckles removed), binarized (converted to black and white), and contrast-enhanced. This stage has an outsized impact on accuracy -- a well-preprocessed image can improve recognition rates by 20-30%.
Layout analysis. The system identifies regions of the page: text blocks, columns, images, tables, headers, footers. This determines the reading order and separates text from non-text elements.
Character segmentation. Individual characters are isolated within each text region. For well-spaced printed text this is straightforward, but for kerned, overlapping, or handwritten text it becomes more challenging.
Character recognition. Each segmented character is classified using pattern matching, feature extraction, or neural network inference. Modern OCR engines use deep learning models trained on millions of text samples across hundreds of fonts and writing styles.
Post-processing. The recognized text is refined using language models, dictionary lookups, and contextual analysis. This catches errors like confusing "rn" with "m" or "l" with "1" -- mistakes that are common at the character level but obvious in context.
Early OCR engines relied on template matching -- comparing character shapes against a library of known fonts. This worked well for clean, standard-font documents but struggled with unusual typefaces, degraded scans, or handwriting.
Modern OCR engines use neural networks, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with attention mechanisms. These models learn to recognize characters from training data rather than hardcoded templates, making them far more robust to variations in font, size, orientation, and quality.
Tesseract, the most widely-used open-source OCR engine, transitioned to an LSTM (Long Short-Term Memory) neural network architecture in version 4. This dramatically improved its accuracy, particularly for non-Latin scripts and degraded documents. ConvertIntoMP4's PDF OCR tool and image to text tool use Tesseract's neural engine with optimized preprocessing.
One of the most important factors in OCR accuracy is language support. OCR engines need language-specific models that understand the character set, common letter combinations, and dictionary of each language.
The following table lists the languages supported by ConvertIntoMP4's OCR engine, along with their script type and typical accuracy range on well-scanned documents:
| Language | Script | ISO Code | Typical Accuracy (Clean Scan) | Notes |
|---|---|---|---|---|
| English | Latin | eng | 98-99% | Highest accuracy, largest training set |
| French | Latin | fra | 97-99% | Excellent with accented characters |
| German | Latin | deu | 97-99% | Handles umlauts and eszett well |
| Spanish | Latin | spa | 97-99% | Good with tildes and accented vowels |
| Portuguese | Latin | por | 97-98% | Includes Brazilian variant |
| Italian | Latin | ita | 97-99% | Strong diacritical mark support |
| Dutch | Latin | nld | 97-98% | Handles digraphs (ij) correctly |
| Polish | Latin | pol | 96-98% | Good with special characters (ł, ą, ę) |
| Russian | Cyrillic | rus | 96-98% | Full Cyrillic character set |
| Ukrainian | Cyrillic | ukr | 95-97% | Distinct from Russian model |
| Arabic | Arabic | ara | 90-95% | Right-to-left, connected script |
| Chinese (Simplified) | CJK | chi_sim | 92-96% | Large character set, context-dependent |
| Chinese (Traditional) | CJK | chi_tra | 91-95% | More complex characters than simplified |
| Japanese | CJK + Kana | jpn | 91-95% | Mixed script (kanji, hiragana, katakana) |
| Korean | Hangul | kor | 93-96% | Syllabic block structure |
| Hindi | Devanagari | hin | 90-94% | Connected headline (shirorekha) |
| Turkish | Latin | tur | 96-98% | Handles dotted/undotted I distinction |
Pro Tip: When scanning multilingual documents, specify the primary language for best results. If a document contains text in multiple languages (for example, an English document with French quotations), select the dominant language. The OCR engine will still recognize common Latin-script words from other languages, though specialized characters may have slightly lower accuracy.
Before running OCR, evaluate the quality of your source material. The single biggest factor in OCR accuracy is input quality.
Good candidates for OCR:
Challenging inputs:
If your source document needs improvement, preprocessing makes a significant difference:
Navigate to the PDF OCR tool on ConvertIntoMP4. Upload your scanned PDF or image file. The tool accepts PDF, JPEG, PNG, TIFF, and BMP inputs.
Configure the OCR settings:
After OCR processing completes, review the output for errors. Common issues to look for:
For critical documents (legal, medical, financial), always proofread the OCR output against the original.
Pro Tip: If you need to OCR a large batch of scanned documents, process a sample of 5-10 pages first to evaluate quality. This lets you identify systematic issues (wrong language model, preprocessing needed, consistent misrecognitions) before processing the entire batch.
Accuracy varies significantly depending on the type of document being processed. This table provides realistic expectations.
| Document Type | Expected Accuracy | Common Issues | Recommendations |
|---|---|---|---|
| Modern printed documents | 98-99% | Very few errors on clean scans | 300 DPI, auto-deskew |
| Business letters and memos | 97-99% | Letterhead logos may confuse layout | Crop headers if problematic |
| Books and publications | 96-98% | Footnotes, page numbers, columns | Specify reading order |
| Legal contracts | 95-98% | Dense text, small font, numbered clauses | High resolution (600 DPI) |
| Historical documents (pre-1960) | 85-95% | Faded ink, yellowed paper, old typefaces | High contrast preprocessing |
| Handwritten text (neat) | 70-85% | Inconsistent letterforms, connected strokes | Moderate expectations |
| Handwritten text (cursive) | 50-75% | Highly variable, overlap between characters | Manual review essential |
| Receipts and invoices | 90-96% | Thermal paper fade, small fonts, mixed layouts | Scan promptly before fading |
| Medical records | 90-95% | Specialized terminology, abbreviations | Medical dictionary post-processing |
| Engineering drawings | 60-80% | Mixed text/graphics, rotated labels, annotations | Extract text regions manually |
| Photographs of documents | 85-95% | Perspective distortion, uneven lighting, shadows | Flatten and correct perspective first |
| Faxed documents | 80-92% | Low resolution, compression artifacts, moiré | Rescan original if available |
If your documents contain specialized terminology (medical, legal, technical), a custom dictionary can dramatically improve accuracy. The OCR engine's post-processing step uses dictionary lookup to correct ambiguous recognitions. When the dictionary includes domain-specific terms, words that would otherwise be "corrected" to common alternatives are preserved correctly.
For example, a medical OCR dictionary would include terms like "brachytherapy," "angioplasty," and "metformin" that a general dictionary might try to "correct" into more common words.
For particularly challenging documents, a multi-pass approach can improve results:
This technique is most useful for historical or degraded documents where a single set of preprocessing parameters does not work well for the entire page.
Tables are one of the most challenging elements for OCR because the engine must understand both the text content and the spatial relationships between cells. Tips for better table OCR:
If you need to convert entire PDF documents while preserving layout including tables, the PDF converter can help maintain document structure.
Scanned PDFs are the most common OCR use case. These are PDFs where each page is essentially a full-page image, usually created by a document scanner or multifunction printer. The PDF OCR tool accepts these directly and produces a searchable PDF with an invisible text layer overlaid on the original page images.
The searchable PDF approach is particularly useful because it preserves the visual appearance of the original document while making the text selectable, searchable, and copy-able. This is the standard approach used by law firms, government agencies, and corporate archives.
Photos of documents -- taken with a smartphone camera, for example -- present additional challenges compared to scans:
For best results with photographed documents, use the phone's document scanning mode (available in most camera apps), which automatically corrects perspective, enhances contrast, and crops to the document edges. The image to text tool processes JPEG and PNG images.
For large document digitization projects, consider this workflow:
Pro Tip: Keep your original scans even after OCR processing. OCR technology continues to improve, and you may want to re-process older documents with newer, more accurate engines in the future. Storage is cheap -- reprocessing is time-consuming.
Documents that contain text in multiple languages (for example, a scientific paper with English body text and Japanese citations) are harder to OCR because the engine optimizes for a single language model.
Solution: Process the document in the dominant language first, then re-process sections in other languages separately. Merge the results manually. Some advanced OCR engines support multi-language detection, but accuracy is typically lower than single-language processing.
Old, faded, or damaged documents produce poor scans that degrade OCR accuracy significantly.
Solution: Apply aggressive preprocessing -- high-contrast binarization, strong noise reduction, and careful deskewing. If the original document is available, rescan at the highest possible resolution. For archival documents that cannot be rescanned, accept that accuracy may be 80-90% and plan for manual correction.
Handwritten text remains one of the hardest problems in OCR. Even neat handwriting produces dramatically lower accuracy than printed text.
Solution: For critical handwritten documents, consider specialized handwriting recognition services rather than general-purpose OCR. For informal needs (handwritten notes, whiteboard captures), general OCR provides a useful starting point that you can then manually correct.
Multi-column layouts, sidebars, callout boxes, and text that wraps around images can confuse OCR layout analysis, causing text from different sections to be mixed together.
Solution: If layout analysis fails, crop the document into simpler regions and OCR each region separately. For consistently formatted documents (like a magazine template), establish a preprocessing workflow that isolates text regions reliably.
OCR is often one step in a larger document processing workflow. After extracting text from a scanned document, you might need to:
The document converter category page provides an overview of all available document conversion tools.
When processing documents that contain sensitive information (financial records, medical data, legal contracts, personal identification), security matters:
Pro Tip: For documents containing personally identifiable information (PII), social security numbers, or financial data, consider processing them in batches and immediately downloading the results rather than leaving files in any cloud service queue. Delete your browser cache after downloading sensitive OCR results.
Law firms process enormous volumes of scanned documents for e-discovery, contract review, and case preparation. Key considerations for legal OCR:
Healthcare organizations digitize patient records for electronic health record (EHR) systems. Medical OCR presents unique challenges:
Banks, accounting firms, and finance departments process invoices, receipts, bank statements, and tax documents:
Before considering an OCR job complete, verify the following:
OCR transforms static document images into dynamic, searchable, editable text -- unlocking the content trapped in scanned pages, photographs, and legacy PDFs. The technology has matured significantly with neural network-based engines, but the quality of your results still depends heavily on input preparation, language selection, and post-processing verification.
For quick one-off OCR tasks, the image to text tool and PDF OCR tool provide fast, accurate results directly in your browser. For large-scale digitization projects, combining good scanning practices with batch OCR processing and systematic quality checks produces the best outcomes.
The key takeaway: OCR accuracy is not just about the engine -- it is about the entire pipeline from scanning to post-processing. Invest time in preparing your source documents and reviewing the output, and you will get dramatically better results regardless of which OCR tool you use.
Michael Rodriguez
Video production expert covering codec standards, streaming formats, and professional post-production pipelines.
