A new method is proposed to select optimal initial cluster centers. By searching parameters' histogram peak values, the number of cluster centers can be estimated, and these optimal initial cluster centers are selected in the columns or cells where the histogram peaks exist. Because these optimal initial cluster centers are near to real cluster centers, the iterations of clustering are reduced efficiently. Theoretical analysis demonstrates that the compute complexity of new method is lower than some conventional techniques. The improved K-Means algorithm is applied to sort frequency-hopping signals, and the simulation results demonstrate that the algorithm is effective.