Abstract: Aimed at the shortcomings of existing models in identifying detailed information, Gabor transform domain Gaussian model vector embedding face recognition method is proposed, called (GGVE). GGVE established a multivariate Gaussian model in the multi-scale transform domain of Gabor wavelet, and used the logarithmic-Euclidean vector embedding method to transform the Gaussian model into a linear space, and used the Euclidean distance to quickly calculate the similarity of the model. Compared with existing hand-crafted descriptors, GGVE could more effectively extract robust facial detail features in complex environments. In order to make up for the information loss of deep network, a multi-feature output ResNet50 network model (ResNet50MF) was proposed, which was combined with GGVE features for face recognition. Experiments show that the fusion of GGVE features and ResNet50MF high-level features can significantly improve the recognition accuracy, and can be applied to face recognition in complex environments.