Adaptive Feature Extraction Method for Degraded Character Recognition

Most character recognition applications target machine printed and handwritten characters on paper documents. Recently, the recognition of text in videos, web documents, and natural scenes has become an urgent demand; research has intensified because this task is difficult to realize (Antonacopoulos & Hu, 2004; Doermann et al., 2003; Kise & Doermann, 2007; Lienhart & Wernicke, 2002; Lyu et al., 2005; Zhang & Kasturi, 2008). The problems posed by recognizing low quality characters in the above mentioned applications are mainly due to deformation such as the variety of font styles and style effects, as well as image degradation like background noise, blur, and low resolution. A key weakness of most conventional character recognition methods is that they tackle either one problem or the other, not both. For overcoming image degradation, some methods, e.g. (Ho, 1998; Kopec, 1997; Xu & Nagy, 1999), design templates that reflect the degradation type anticipated. Also a robust discriminant function for recognizing degraded characters was proposed in (Sato, 2000; Sawaki & Hagita, 1998). Unfortunately, these methods are sensitive to shape deformation, since they employ image-based template matching. They fail to effectively handle multiple fonts and several style effects. On the other hand, geometric features are often used for recognizing multiple fonts. Stroke direction is particularly effective against character deformation (Umeda, 1996). For example, the direction contribution based on stroke run-length is effective (Akiyama & Hagita, 1990; Srihari et al., 1997; Zhu et al., 1997). However, geometric features are not robust against corruption of information due to image degradation. In addition, although geometric features are more robust against deformation than image-based template matching, they are not invariant for deformation such as aspect ratio fluctuation and stroke position shift. Therefore, geometric features are weak against the kinds of deformation that are not present in the training samples. For overcoming deformation problems mentioned above, nonlinear shape normalized techniques (Tsukumo& Tanaka, 1988; Yamada et al., 1990) have been proposed as a pre-processing method to relocate strokes uniformly. They normalize a pattern by exploiting the distance between strokes (Tsukumo & Tanaka, 1988) and stroke line density (Yamada et al., 1990), and are mainly aimed at the recognition of Kanji characters that consist of many strokes in mostly square patterns. Therefore, applying these methods to the recognition of numerals, alphabets and kana characters, which consist of fewer strokes and are not square shape, is difficult. Also these methods are ineffective for degraded characters with backgrounds noise and blur be3