| Document analysis and recognition is concerned with the automatic interpretation of images of printed and handwritten documents. The recognition of human handwriting, including numerals, characters and signatures, has long been an important research topic in handwritten document interpretation. The prime difficulty in handwriting identification and recognition lies in a variety of deformation of character shapes. In order to break through the difficulty, it is essential to study primary and discriminating characteristics of handwriting and select a set of features which are both sufficiently invariant to tolerate the great variability of handwriting and sufficiently significant to determine which word is written. In this project, the intrinsic factors of handwriting are studied from the biophysical and psychophysical points of views and a set of geometric and topologic features are chosen to represent a cursive script. These features are discriminating, robust and perceptually salient from biophysical and psychophysical viewpoints, and describe the intrinsic factors of shapes and curves that make up a cursive handwriting. From the graphological point of view, these features can be classified into two groups: regular and singular. Regular features include geometric shapes such as horizontal bars, vertical bars and loops. Singular features include singularities and quasi-topologic features such as end points, branch points, crossing points, convexity, and concavity. Incorporating the proposed new handwriting representation scheme, an off-line signature identification system, an off-line signature verification system, an off-line cursive handwriting segmentation system, and an off-line handwritten word recognition system are constructed and implemented in this work. Since the features in the proposed new handwriting representation scheme are carefully selected based on biophysical and psychophysical processes involved in handwriting generation and represent global/local, structural/statistical and geometric/topologic information of handwriting, all the proposed systems achieve good performances and provide high accuracy rates. |
