Generative Adversarial Network Based Entangled Face Representation Analysis

Face recognition is one of the most important biometric recognition technologies,and serves in many areas like video surveillance and identity verification.The goal of a recognition system is to extract facial identity information from an input picture.However,every people may have hundreds or thousands pictures,and these pictures differs from each other in poses,lights,surroundings and expressions.The influence is shown in two aspects:first,the wanted identity information is fully entangled with variance,and this makes the system fail to search the true identity;second,in many real world scenarios,it's really difficult to collect efficient pictures for every subject,and this do not guarantee to learn a robust system because of the lack of intra variance.For the above two challenges,this paper leverages Generative Adversarial Network technology,and introduces face representation learning methods to represent both identity and variance information.The methods are performed on pose-invariant Face recognition and one-shot face recognition problems.For pose-invariant face recognition,we propose an improved method called Cross-generating,and we are the first to automatically encoding both identity and non-identity in two separate vectors,thus successfully preserve constant identity information from posed pictures;we also leverage a Generative Adversarial Network to in training phase,and this operation helps the model reconstruct clear and real multi-pose face pictures.Comparative experiments are performed on MultiPIE dataset,and our method achieves the state-of-the-art recognition accuracies.For one-shot face recognition,we propose to augment the feature space of a subject by learning to transfer within its class.We start from the representation transfer within the subject,build a new network block into a usual classification network,and the block performs inner transfer by adding a random learnable variable to an intermediate feature;by leveraging adversarial network,the model learns the transfer mode in an end-to-end manner.The comparative experiments on MS-Celeb-1M Challenge-2 show that the coverage@precision=99%is largely raised from 26.77%to 91.41%,thus the extensiveness of a recognition system is improved.