Incorporating contextual character geometry in word recognition

Abstract

Contextual character geometry is the geometric information available only when a character presents in the context of a word. Such information includes the character's location and relative size in the entire word image, forming a bounding box of the character. The differences between the geometry of an image segment and the expected geometry of a candidate character are considered as additional features to refine the recognition of individual characters. A typical word recognizer based on over-segmentation and segment-combination is used to illustrate the use of these new features and experimental results have shown significant improvement of recognition accuracy, especially on large lexicons.