Recognition of handwritten symbols

Abstract

New statistical methods are presented for recognizing isolated, handwritten or printed symbols directly from raster images of documents such as technical drawings and maps. Our methods for recognizing symbols of known or unknown rotation avoid the traditional thinning and vectorization steps in the recognition process. Hence, the information loss caused by these two steps is eliminated. In all of our approaches we consider the outer pixel boundary of an isolated symbol candidate in the binary raster image as a simple closed curve. This curve is then approximated by a parametric spline curve, an elliptic Fourier expansion of the Fourier expansion due to Zahn and Roskies. Curvature values and coordinates along the spline curves or the coefficients of the Fourier expansion are then used as descriptors in a statistical classification scheme. Statistical classification using these methods was done on separate training and test sets consisting of the digits 0 to 9 and all lower-case letters. There were about one hundred samples of each handwritten symbol class in each set. The correct classification rate was as high as 97.7% using the parametric spline approximation, and 98.6% using the elliptic Fourier descriptors.