Optical Character Recognition
The conversion of images of text — scans, photographs, screenshots — into machine-readable text.
Optical character recognition (OCR) converts images of text — scanned documents, photographs, screenshots — into machine-readable text. The classical pipeline consists of preprocessing (binarization, deskewing, denoising), segmentation into lines, words, and characters, feature extraction, and classification by a model trained on labeled character images. Tesseract is the best-known open-source engine of this tradition.
Traditional OCR relied on template matching and hand-crafted features, which made it sensitive to fonts and noise. Modern systems are deep neural networks: convolutional neural networks classify characters robustly, and sequence models — recurrent networks with connectionist temporal classification (CTC), and more recently transformer-based architectures — recognize whole lines of text directly from images without explicit character segmentation. End-to-end systems handle multiple scripts, fonts, and mixed layouts.
Challenges remain: degraded or handwritten documents, skewed or perspective-distorted images, low resolution, and non-Latin scripts. Systems are evaluated by character and word error rates, and state-of-the-art results combine OCR with layout analysis and document understanding to extract structured data from invoices, forms, and books.
The applications are ubiquitous: digitizing books and archives, automatic number-plate recognition, bank-check processing, accessibility tools that read printed text aloud, and real-time translation of photographed signs. OCR is one of the oldest and most commercially successful uses of computer vision, and deep learning has improved its accuracy dramatically.
Tags
computer vision document processing machine learning ocr
Related articles
- k-Nearest Neighbors
- k-Means Clustering
- Cross-Validation
- Word Embeddings
- Gradient Descent
- Overfitting and Regularization
Click here for easy-to-read helpful e-books for anyone, anywhere, and about anything