Improving pose manifold and virtual images using bidirectional neural networks in face recognition using single image per person

In this article, for the purpose of improving neural network models applied in face recognition using single image per person, a bidirectional neural network inspired of neocortex functional model is presented. In the proposed model, recognition is not performed in a single stage, but via two bottom-up and top-down phases and the recognition results of first stage is used for model adaptation. We have applied this novel adapting model in combination with clustering person and pose information technique to separate person and pose information and to estimate corresponding manifolds. To increase the number of training samples in the classifier neural network, virtual views of frontal images in the test dataset are synthesized using estimated manifolds. Training classifier network via virtual images obtained from bidirectional network, gives an accuracy rate of 86.36% on the test dataset which shows 15.46% improvement in accuracy of face recognition compared to training classifier with only frontal view images.