An online handwritten numerals segmentation algorithm based on spectral clustering

Abstract

In our previous work, without considering the stroke information, a method based on spectral clustering (SC) for solving handwritten touching numerals segmentation was proposed and obtained very good performance. In this paper, we extend the algorithm to an online system, and propose an improved method where the stroke information is involved. First, the features of the numerals image are extracted by a sliding window. Second, the obtained feature vectors are trained by support vector machine to generate an affinity matrix. Thereafter, the stroke information of original images is used to generate another affinity matrix. Finally, these two affinity matrices are added and trained by SC. Experimental results show that the proposed method can further improve the accuracy of segmentation.