A Survey on Optical Character Recognition System

I. INTRODUCTION

• OCR digitizes printed text and images into a form that machines can manipulate.Machines do not perceive information in images as readily as humans.Source: §I · p. 1
• OCR is challenging because languages vary in their rules, and text appears in many languages, fonts, and styles.The paper identifies image processing, pattern classification, and natural language processing as disciplines used to address OCR challenges.Source: §I · p. 1
• The paper surveys OCR history, applications, challenges, and techniques, including proposals addressing OCR issues.Source: §I · p. 1

II. LITERATURE REVIEW

• The earliest systems were slow and inaccurate; GISMO read 23 characters, while Rainbow read uppercase typewritten English at one character per minute.GISMO, invented by M. Sheppard in 1951, could also read musical notation.Source: §II · p. 1
• OCRA and OCRB were developed in 1970 to standardize fonts and ease recognition, with comparatively acceptable recognition rates.Research later produced document image analysis and multilingual, handwritten, and omni-font OCR.Source: §II · p. 1
• Despite extensive research, machine reading remains below human ability, and OCR for complex languages such as Urdu and Sindhi has lacked available software.Current research targets accuracy and speed for unconstrained printed and written documents.Source: §II · p. 1

III. TYPES OF OPTICAL CHARACTER

• Character-recognition systems are categorized by image acquisition mode, character connectivity, and font restrictions.Figure 1 presents the categorization.Source: §III · pp. 1–2
• Handwriting recognition is difficult because writing styles and pen movements vary, while machine-printed recognition is relatively simpler when character dimensions and positions are predictable.Source: §III · pp. 1–2
• Online handwriting systems capture stroke timing and direction during writing, whereas offline systems recognize static bitmaps and are described as difficult.Online systems can use temporal information and do not require thinning; many have been developed for tablets and PDAs.Source: §III · p. 2
• OCR applications include mail sorting, cheque reading, automated form processing, passport validation, number-plate recognition, and helping blind and visually impaired people read [6][7].Source: IV.APPLICATIONS OF OCR · p. 2

IV. MAJOR PHASES OF OCR

• Segmentation separates characters before recognition and may be explicit or implicit in classification; overlapping, broken, or noisy characters need advanced techniques.Connected-component analysis and projection profiles are among the simplest techniques.Source: §IV · pp. 2–3
• OCR comprises acquisition, preprocessing, segmentation, feature extraction, classification, and post-processing rather than one atomic operation.Overlapping or broken characters and noise can require advanced segmentation; post-processing may use language, geometric, and linguistic context to correct errors.Source: §IV · p. 2
• Post-processing can improve OCR accuracy by using natural-language, geometric, and linguistic context, including spell-checking, dictionaries, Markov chains, and n-grams.It should avoid excessive time and space costs and should not introduce new errors.Source: §IV · pp. 2–3
• Image acquisition obtains a digital image and converts it into a computer-processable form, potentially using quantization, binarization, and lossy or lossless compression.Source: §IV · p. 2
• Preprocessing enhances image quality through thresholding, filtering, skew estimation, optional thinning, and text-line detection.Thresholding may use a local or global value; line detection may use projections or pixel clustering.Source: §IV · pp. 2–3
• Feature extraction selects character representations such as images, geometrical features, or statistical moments, with principal component analysis available for dimensionality reduction.Features should be efficiently computable and reduce within-class variation while increasing between-class variation.Source: §IV · p. 3
• Classification maps character features to categories using structural, statistical, or syntactic approaches, including Bayesian, decision-tree, neural-network, and nearest-neighbor classifiers.Source: §IV · p. 3
• Multiple classifiers can be combined in cascading, parallel, or hierarchical arrangements to improve OCR results.Their outputs can then be combined using different approaches.Source: §IV · p. 3

• OCR comprises acquisition, preprocessing, segmentation, feature extraction, classification, and post-processing, whose techniques the paper reviews.Source: §V · p. 3
• Combining OCR techniques is proposed as future work for developing efficient systems, with number-plate recognition, smart libraries, and other real-time applications identified as uses.Source: §V · p. 3
• Recognition for Arabic, Sindhi, and Urdu remains an open challenge, and an overview of techniques for these languages is planned as future work.Source: §V · p. 3
• Multilingual character recognition and practical OCR applications remain active research areas.Source: §V · p. 3