The simplicity and efficiency of the community detection algorithm based on label propagation (LPA) have been studied extensively, but when the community structure is not clear, a single community is obtained through the LPA, which is meaningless. Modularity-specialized label propagation algorithm (LPAm) tends to partition the network into communities with similar degrees and the problems of solving the limit of functions exist. Therefore, this paper points out label propagation algorithm based on modularity density optimization (LPAd), in order to avoid the formation of large communities, and the community meets the weak community definition proposed by Radicchi et al. Several real datasets and artificial network data experimental results show that, this algorithm raises the quality of the detected community without changing the algorithm complexity, and compared with the existing number of community detection algorithm based on label propagation, it has been improved effectively.