PRACTICAL BASIS SET FOR CHINESE CHARACTER REPRESENTATION.

Abstract

Chinese characters have the property that the strokes are distributed mainly in the horizontal, vertical, and slant directions. Hence, there are strong correlations in these directions that can be utilized for data compression. Among the decorrelation techniques, the Karhunen-Loeve transformation (KLT) is optimal. However, it is not practical due to the heavy computation involved in diagonalizing the N**2 multiplied by N**2 correlation matrix corresponding to the sample images digitized on an N multiplied by N grid. A modified KLT (MKLT) that can be efficiently computed is proposed. The method is to transform the character with respect to several basis sets, each set representing strokes of a given direction. It is shown that by suitable modeling, computation of the MKLT involves only diagonalization of N multiplied by N matrices instead of N**2 multiplied by N**2 matrices as generally required. The inherent computational problem of the KLT is thus bypassed by this suboptimal approach. Experimental results show that the MKLT is more effective than Fourier transformation for Chinese character representation.