An improved spectral clustering algorithm based on low rank approximation for image segmentation

Spectral clustering algorithm requires huge storage and has high computational complexity. It is hard to be applied in large image processing. An Improved Spectral Clustering based on Low Rank Approximation (ISCLRA) is proposed in this paper. It can solve the eigen value decomposition (EVD) of the large scale matrix with lower computational complexity. Next, by incorporating the Mean Shift (MS) and ISCLRA, a novel color image segmentation method named MS-ISCLRA is designed in this paper. First, MS-ISCLRA preprocesses the input image by using the MS to form segmented regions. Second, it regards the average color vectors of the segmented regions as the inputs of the ISCLRA. Finally, it applies the ISCLRA to perform globally optimized clustering. Because the number of the segmented regions is much smaller than that of the image pixels, MS-ISCLRA allows color image segmentation with significant reduction of the complexity. The experiments illustrate that MS-ISCLRA has superior performance in image segmentation than MS-Ncut.