A system and method of mode selection face recognition with parallel CNNs are disclosed. The system includes a mode selection module, a parallel convolutional neural network module, a database module and an ID judgment module. The parallel convolutional neural network module includes several convolutional neural network units. The mode selection module can determine the facial characteristics of each facial image, and send each facial image to the convolutional nerve corresponding to its facial characteristics The results serve as a benchmark for the ID judgment module to compare and test other facial images. By using several convolutional neural network units for facial image training for different facial characteristics of the same person, the noise from other convolutional neural network units can be reduced and the similarity in the feature vector extraction stage can be improved. The resulting accuracy of face recognition will increase.