Abstract

Intense activity and significant progress have characterized the last decade in the field of the recognition of unconstrained handwritten numerals by computer. The diversity and ingenuity of the methods proposed are carefully reviewed at the beginning of this thesis and the results achieved are compared. Despite important advances, the very high reliability of human recognition has not been matched by these approaches and our work is intended as a contribution towards bridging this reliability gap. In recent years, the combination of several recognition methods has been a very fruitful idea in this regard. Here we explore another avenue: overcoming the limits of single methods, particularly structural model-based methods, to deliver much more reliable classification on their own. Lessons are drawn from past work and all stages of the recognition process which are typical of this approach (preprocessing, feature extraction, and classification) are revisited. Achieving higher levels of reliability and robustness in the feature extraction stage was seen as key to achieving our goal. Hence much of our research was devoted to the solution of this problem including a comparative study of several curvature feature extraction schemes, the detailed 'autopsy'of a feature extractor previously devised by this author, and the meticulous construction of a new extractor to circumvent identified weaknesses. Compared to a preceding effort at creating a numeral recognition system, our conception of the development of classification rules was also deeply revised. Much care is taken to distinguish all the (global or local) shape variants to be identified by the system and to tightly model each of these variants in a more refined and exhaustive manner. A special syntax and a development interface were designed to assist in this task. Results for the CENPARMI database, from a partial classifier built upon these foundations, demonstrate the feasibility of creating a single-method numeral recognition system with a high recognition rate (around 90%) and a very low substitution rate. When applied to other databases, including some which incorporate markedly different writing styles, the very high reliability of the system is maintained