A Research On Human Face Multi-Feature Fusion Related Technologies Based On Frame Sequence

In recent years,deep learning technology has greatly promoted technological innovation in the field of face recognition,and the recognition rate on various public data sets has made great breakthroughs.However,in some natural scenes of the practical applications,the face recognition effect is not ideal due to problems such as low resolution and unbalanced illumination.At the same time,most of the current face recognition technology is applied to video,but only uses single frame data in the video for identification,which causes a lot of information waste.In summary,this thesis proposes a face recognition technology based on frame sequence,which can fuse the features of different frames and blocks in the same frame to achieve multi-feature fusion.The details are as follows:1.Aiming at the problem of super-resolution,this thesis proposes a new superresolution network based on interpolation generative adversarial network(SRIGAN).According to proposed different interpolation methods,SRIGAN can be divided into two networks: super-resolution self-interpolation generative adversarial network(SRSIGAN)and super-resolution channel interpolation generative adversarial network(SRCIGAN).In the experiment,this thesis analyzes the effects of different residual block numbers and activation functions in the two networks.Then,a comparison experiment is carried out from two aspects: the convergence speed on the training set and the resolution improvement effect on the test set.The results show that both types of models can achieve faster convergence speed while maintaining excellent test results.2.Aiming at the multi-feature fusion problem of face,this thesis proposes three feature fusion models BA-FA,FA-FA and BA-FA-FA based on attention mechanism,which accomplish fusion of frame block features and frame features on different layers.After adjusting the appropriate parameters through experiments,further experiments are carried out.The results show that both the common feature extraction network and the advanced feature extraction network,combined with the three feature fusion models proposed by this thesis,can be effectively improved on the final feature identification ability.The test results based on VGG-19-128 show,the average AUC values of the three feature fusion models increased by 7.62% and the maximum increased by 10.06%.3.For the face verification problem under low resolution,combined with the superresolution network and feature fusion model mentioned above,this thesis proposes three networks: BF-GFN,FF-GFN and BFF-GFN.The test on the monitoring face data set of the low resolution provided in this thesis has achieved better results than the mainstream method.Finally,based on the work of this thesis,the face verification system is implemented.And from the actual situation,the feasibility of our models is confirmed.