Phase unwrapping plays an important role in three-dimensional (3D) measurement technologies, and its analytical accuracy directly affects the accuracy of 3D modeling. Due to undersampling and discontinuity of the wrapped phase, it is difficult to obtain correct phase information for traditional spatial phase unwrapping, while temporal phase unwrapping requires additional auxiliary information. For 3D face modeling in complex scenarios, a phase unwrapping network based on multi-scale attention is proposed in this paper. In this network, the encoder-decoder structure is used to fuse multi-scale features, and an attention sub-network is embedded into the decoding network for contextual information collection. A FACE dataset of 5000 samples and a MASK dataset of 100 samples are constructed, and each sample contains the truth values of wrapped phases and continuous phases for training and testing of phase unwrapping. The root-mean-square errors of the proposed network are 0.0387 rad and 0.0273 rad on the FACE dataset and the MASK dataset. The structural similarities are 0.9850 and 0.9793 respectively. The phase features can be extracted quickly and accurately in areas such as undersampled and phase discontinuous ones to ensure the correctness of phase unwrapping. Finally, the effectiveness and feasibility of the proposed network are verified by comparative experiments.