A General Approach for Handwritten Digits Segmentation Using Spectral Clustering

In this paper, an approach is proposed to solve a classic segmentation problem on handwritten touching digit pairs using spectral clustering (SC). SC has appeared in many of the state-of-the-art algorithms on image grouping problems recently, while it is a challenging work to build similarities between each pixel. In this paper, support vector machine (SVM) is used to predict the affinity matrix in SC instead of designing a complex function, hence making it a general approach. Different from traditional methods which focus on finding the cutting points or lines, we treat the handwritten string segmentation as a graph partitioning problem, which enables us to separate those digits connected in a very complicated way. We also introduce a 'second-segmentation' to optimize the segmentation result, and find out that the whole algorithm is similar to a multi-layer perception (MLP). Experiment results show that the proposed approach performs satisfactorily with high correct rate while keeping its own advantages as a general approach.