Face verification based on deep reconstruction network

Face verification is a difficult problem for the variations of illumination, expression and pose. Inspired by the denoising auto-encoder, a new deep learning algorithm named deep reconstruction network (DRN) is proposed to improve the face verification accuracy under these variations. In this proposed DRN, the input face containing variations is fed to the encoder and the target of the decoder is to reconstruct its reference version. With this object, the DRN models the input face as the reference face contaminated by the illumination, pose and expression, and thus the extracted facial features are insensitive to the three variations. Experimental results on the CMP-PIE dataset demonstrate that the DRN has good capability to reconstruct the reference face, and the facial features extracted by DRN have good robustness to obtain high face verification accuracy.