| Most current OCR systems perform a binarization of the input before attempting recognition. Though this might appear to be a rather innocuous operation since the input is supposed to be binary to start with, a significant amount of information is lost during the process. To avoid this problem, other systems perform recognition without binarization by using techniques related to matched filters. However, this approach is limited to applications where there is only a limited number of typefaces.; In this thesis, a new approach is explored; this approach eliminates binarization by extracting features directly from gray scale images. A digitized gray scale image is treated as a noisy sampling of the underlying continuous surface and desired features are obtained by extracting and assembling topographic characteristics of this surface. The advantages and effectiveness of the new approach are both shown theoretically and demonstrated through experimental results.; Polygonal approximations are widely used by conventional structural pattern recognition methods. However, they have the disadvantage of introducing significant ambiguities. In particular, vertices may correspond either to real corners or to approximations of smooth arcs. Such ambiguities, if not properly dealt with, may cause misclassifications and thus seriously affect the OCR system performance. To solve this problem, a novel method for feature decomposition is proposed based on the same platform of our gray scale approach without using polygonal approximations. Decomposition is achieved using both tangent discontinuities and curvature primitives, which are computed from gray scale topography. The new method not only eliminates the ambiguities associated with polygonal approximations completely but also helps to solve shape confusions among similar characters. |
