As one of the commonly used data mining algorithms, K-means has the advantage of fast clustering speed, but the disadvantage is that it is less effective for clustering non-spherical data. An improved K-means algorithm (IK-means) is proposed to enhance clustering efficiency for non-spherical data. The original dataset is clustered into a relatively larger number of high-density sub-clusters, and the final result is obtained by merging connected sub-clusters respectively. The connectivity among sub-clusters is evaluated by the sub-clusters density and the nearest distance class between sub-clusters. By testing on University of California, Irvine(UCI) datasets and several other artificial simulation datasets, the comparison of proposed IK-means algorithm against DBSCAN, KGFCM shows its clustering capability for data of arbitrary shape. The clustering Adjusted Rand Index (ARI) value for 72,000 sizes data is 24% higher than DBSCAN, and 95.2% higher than KGFCM. For larger datasets, the IK-means algorithm is faster than DBSCAN and KGFCM.