The face recognition system consists of a face detection module, a face alignment module, a feature extraction module, and a feature matching module. The face detection and alignment modules are modified from existing methods to improve both computational efficiency and accuracy. The facial feature extraction module is developed based on the deep learning technique. We proposed a new extremal feature map (EFM) as the nonlinear activation function and include ResNet blocks to our deep neural network (DNN) which contains only 8.5 million parameters. The DNN is trained with 1.1 million images of about twenty thousand persons to compute a 256-D feature vector. The proposed system is tested using the LFW face recognition benchmark dataset and the overall recognition accuracy is 98%.
Structure of the Proposed DNN for facial feature extraction
nVidia Graphic Card (Compute Capability >= 2.0)
Intel i5 or i7