Abstract: For the low-quality 3D face data collected by the fast collection device, we propose a face recognition method which consists of Spatial Attention-based Dropout (SAD) and Inter-class Regularization Loss function (IR Loss). This method effectively improves the recognition accuracy of incomplete 3D face data. SAD allows the network to learn more hidden features by randomly dropping out the important parts of the feature map based on the spatial attention mechanism. And IR Loss makes the feature similarity between faces with different identities lower by restricting the distance separation of the class centers between them. Experiments show that the method proposed in this paper is superior to the current benchmark methods on a largest low-precision dataset (Lock3DFace), and the SAD and IR Loss we proposed show strong applicability and robustness.