Abstract: Spectral clustering for image segmentation is difficult to calculate the spectrum of weighted matrix. A multistage sampling spectrum image clustering algorithm was designed to eliminate it. First, the sampling theorem was given and proved, and the minimum sample size was derived according to the smallest cluster and cluster number. Secondly,image pixels were sampled based on the minimum sample size and a penalty function was proposed by repeated sampling for the elimination of sampling model on the stability effects of spectral clustering. Finally, the distance between a pixel point and categories was definited, the remaining points were assigned respective cluster depending on the principles ofthe nearest distence. I}he experimental results show the effectiveness of the algorithm.

Key words: Spectral clustering,Multi sampling,Image segmentation,Normalized cut