Most current face recognition algorithms are difficult to deploy to mobile terminals and embedded devices because of their many parameters and large amount of calculation, thus a face recognition method based on improved MobileFaceNet is proposed. By adjusting the structure of the MobileFaceNet model, the bottleneck module is optimized to the sandglass module, the relative position of depthwise convolution and pointwise convolution are improved, and the number of output channels of the sandglass module is appropriately increased, thereby reducing information loss during feature compression and enhancing extraction of the facial spatial features. The experimental results show that the improved method achieves an accuracy of 99.15% on the LFW test data set, and the model size and calculation amount are only 61% and 45% of that of the original algorithm, verifying the effectiveness of the proposed method.