Feeling Information Extraction from Characters

An automated feature extraction and evaluation method of feeling information from printed and handwritten characters is proposed. This method is based on image processing and pattern recognition techniques. First, an input binarized pattern is transformed by a distance transformation. Second, a two-dimensional vector field is composed from the gradient of the distance distribution. Third, a divergence operator extracts source and sink points from the field, and also the vectors on those points. Fourth, the Fourier transform is done for the vector field as a complex valued function. Differently from conventional methods, we deal with the Fourier transform with Laplacian operated phase. Fifth, applying the KL expansion method to the data of the complex vectors obtained from some kinds of character fonts, we extract some common feature vectors of each character font. Using those common vectors and linear multiple regression model, an automated quantitative evaluation system can be constructed. The experimental results show that our vector field method using the combination of Fourier transform and KL expansion is considerably more efficient in the discrimination of printed characters (or fonts), comparing with conventional method using gray level (or binarized) character pattern and KL expansion. Moreover, we obtain the results that the evaluation system based on the regression model comparatively meets well to the human assessment. Key-Words: Feeling information, characters, image processing, vector field, Fourier transform, KL expansion, linear multiple regression model