Recently, the faster R-CNN has demonstrated impressive performance on various object detection benchmarks, and it has attracted extensive research interests. We train a faster R-CNN model on the WIDER face dataset with the ZF and VGG16 convolutional neural network respectively, and then we test the trained model on the FDDB face benchmark. Experimental results demonstrate that the method is robust to complex illumination, partial occlusions and facial pose variations. It achieves excellent performance in detecting unconstrained faces. The two kinds of network have their own advantages in detection accuracy and efficiency, so we can choose to use an appropriate network model according to the actual application requirements.