| dc.description.abstract | Computer Recognition of Handprinted Characters is a well-established field of research, active for over three decades. Despite extensive work and numerous publications, the ultimate goal of developing a recognizer with human-like reading capabilities remains unachieved. Further research is needed to incorporate human reading abilities into recognition systems.
The development of a typical recognition system for an alphabet set involves two major components:
1. Learning – This includes collecting character samples, selecting appropriate features, and developing clear prototype class descriptions.
2. Recognizer Design – This involves designing subsystems for character preprocessing, feature extraction, description generation, and recognition analysis.
The central theme of this thesis is the proposal of a feature-descriptive approach and the automation of recognizer design for various handprinted character sets. The focus is on English (Latin) alphanumerals and a representative South Indian cursive (unconnected) character set. The characters are described by independent features that directly reflect the structure of character patterns. A unified and flexible feature representation is used to handle character sets with varying structural complexities. Each feature is defined by attributes such as:
• Its location in the character frame
• The number of limbs diverging from the feature point
• The directions, curvatures, lengths, etc., of these limbs
Prototype class-descriptions are constructed using these stable and reliable features. To accommodate variations in writing styles and sample deviations, tolerances are introduced for the feature components.
Recognition is performed by comparing the unknown character's description with the prototype, following a pre-classification based on feature locations and limb counts. A Class-Feature Table (CFT) is used to facilitate this process.
The description and recognition procedures are automated through:
1. Automating the feature selection problem
2. Automating the learning of prototype descriptions
Feature selection is simplified to choosing the appropriate components of the defined features. This is guided by a criterion called ambiguity of descriptions, ensuring that only the necessary and sufficient components are selected to achieve unambiguous class-descriptions.
Descriptions are learned automatically from a learning sample set using the selected features. These learned descriptions are stored in a Class-Feature Table (CFT) to enable fast pre-classification. The system adapts to future samples by combining new descriptions with previously learned ones.
This automated approach offers a fast and efficient method for designing recognizers for various alphabet sets. The prototype structural description learning technique eliminates the need for manual inspection and feature selection by the designer.
Both manual and automated approaches have been tested on:
• English (Latin) alphanumerals
• Malayalam, a South Indian character set with approximately 45 basic cursive characters and symbols
The system achieves reasonably high recognition rates without relying on contextual information.
The automated method is practical and well-suited for character sets with fewer than 100 characters or symbols, and shows promising results for stroke-intensive characters. Its extension to larger cursive character sets may be feasible and is worth further exploration.
The thesis is organized into 8 chapters, with Chapter 1 providing an introduction and overview of the stages involved in recognizer development.
The thesis is organized into eight chapters, each focusing on different aspects of character recognition systems and related technologies:
• Chapter 1 introduces the character recognition system, its applications, the problem statement, and motivations.
• Chapter 2 reviews the state-of-the-art in character recognition research, covering major methodologies and recent work. It also briefly discusses practical OCR systems available in the market and their general performance.
• Chapter 3 describes the generation of artificial character databases used throughout the thesis.
• Chapter 4 presents a new adaptive thinning algorithm, designed to minimize or eliminate noise typically introduced by conventional thinning methods.
• Chapter 5 introduces a fast segmentation algorithm for detecting feature points using curvature analysis and neighborhood window examination.
• Chapter 6 covers feature representation, description generation, and recognition techniques.
• Chapter 7 details sequential feature selection and automated description generation, along with recognition performance results based on implementations in Pascal on the BBC 1090 system.
• Chapter 8 concludes the thesis and offers suggestions for future work.
The thesis also includes an extensive bibliography with 235 references and four appendices. | |
