Regenerative neural network for face recognition in video

In this article, the neural network (RNAN) for facial recognition in video is presented. This network receives a sequence of face frames in the video as input, and after repairing and re-creating the frames, it forms a compact representation with fixed dimensions of the features of the frames. The proposed RNAN network has three basic parts. The first part of the network receives the frames and uses the adversarial generator network to recreate the face or restore the quality of low-quality frames. The second part of the network is a residual network (ResNet) which is used to extract features from frames. Finally, the third part receives a number of features extracted from the frames and produces an aggregated unit vector as output. This output is used for identity verification and face recognition in the video. The efficiency of the two end parts of the network is compared on the IJB-A data set and the final results are expressed on the presented TV-Dataset data set. The results show that the RNAN network visibly performs better than simple aggregation networks.

کلیدواژه ها:

Face recognition in video ، neural aggregation network ، restorative network ، reconstructive network ، adversarial generative network ، attention mechanism