Robust Character Segmentation and Recognition Schemes for Multilingual Indian Document Images

This paper presents robust methods for character segmentation and recognition for multilingual (Latin and Devanagari) Indian document images. The documents degraded over the years because of text diffusion, layout structures, and intermixed texts (printed and handwritten). In character segmentation algorithm, initially, pre-processing like noise and non-uniform illumination reduction is performed. Segmentation columns between characters are obtained using their structural properties. In addition, graph distance theory is used to separate touched and overlapped characters. Finally, support vector machine classifier is employed to validate the segmentation results. For character recognition, set of features based on geometrical properties of the characters are proposed. First feature is formed using center pixel of each non-overlapping block, whereas second feature is a vector of cut points calculated in the triangular areas of the character. Apart from these features, one more feature is used that contains neighborhood information of text pixels. Comprehensive experiments are performed on publicly available databases along with the database collected by us. Benchmarking analysis shows that highest segmentation and recognition accuracies are obtained 98.79% and 99.6%, respectively.