Label Propagation Algorithm (LPA) is a fast and efficient community detection algorithm and this algorithm does not need to know the prior information such as the number of communities. However, this algorithm has a large number of randomness, which leads to unstable results. In order to improve the stability of label propagation algorithm, we propose an improved label propagation algorithm (LPAMP). The algorithm is divided into two phases. In the first phase, vertices are clustered roughly by optimizing the modularity greedily; in the second phase, the labels propagate through the network based on the result of the first phase. Experimental results show that the proposed algorithm not only reduces the randomness of the label propagation algorithm, but also improves the stability and increases the accuracy.