Abstract: For the shortage of K-means，a new K-means algorithm is proposed to oprimize the initial center.Firstly，density-sensitive similarity measure is used to compute the density of objects.Based on the distance of objects and the neighborhood of object，the set of high density is obtained，and from which select data points whose mutual spearation is the greatest as possible as they can as initial centers.Then，a rough set-based K-means algorithm is uesd to deal with boundary region，and getting to the cluster number automatically by means of equalization funtion.Experimens show that the method has better cluster results and general performance.