Improved text image segmentation based on spectral clustering

This paper proposes an improved spectral clustering method for image segmentation based on normalized cut (Ncut). In order to effectively reduce the computational complexity of spectral clustering, the method uses color sets quantized as vertexes of graphs to simplify the weighted graph model. Firstly, the similarity function is established according to the characteristics of text images. And then, the color space is quantified by using the color histogram according to the color distribution of scene images, and the affinity matrix is constructed under the quantized levels. Finally the method uses the spectral clustering to segment images under the Ncut criterion. The experiments conducted with a large number of scene images including a publicly available database from the contest of ICDAR 2009 and 2003 show that the proposed method has the good performance in text image segmentation.