When current clustering algorithm is applied to large-scale and numerous-class data sets, the computational complexity is large, and the performance of the algorithm is seriously dependent on the value of K. An improved accelerated K mean clustering algorithm is proposed, which is mainly composed of two strategies, one is Pivot Lower Bound(PLB) based skipping. The fixed points called pivots are newly introduced to calculate the lower bound on the distance between an object and a centroid. PLB avoids unnecessary distance calculations by using the lower bounds, especially early in the convergence process of common clustering algorithm. The other is the Invariant Centroid Pair(ICP) skipping which keeps an object assignment without the distance calculation to an unassigned centroid if the assigned and the unassigned centroids do not change their positions after the centroid update step. In addition, an extended algorithm which combines the proposed algorithm with Hamerly algorithm is presented to further improve the clustering acceleration effect. The simulation experiments of large scale high dimensional image data sets show that, compared with Hamerly algorithm, the proposed algorithm can greatly reduce the number of distance calculations compared with Lloyd algorithm, while realizing the same clustering result. When the K value is large, the proposed algorithm achieves a higher average reduction rate and a shorter average elapsed time.