Research On Unconstrained Environment Face Recognition Technology

In recent years,more and more attentions have been paid on face recognition technology and set of the climax of application of face recognition,including finance,social security,education,intelligent security system and so on.At present,face recognition has achieved promising results on constrained environment,however it is still a challenging problem on unconstrained environment due to the variation of illuminations,expressions,occlusions and so on.These factors produce varying degree of influence on face recognition pipeline,including face pre-process,face feature extraction and face classification,and lead to unsatisfied face recognition accuracies.To deal with these problems,this dissertation mainly focuses on each stage of face recognition system.The basic theories and methods of unconstrained face recognition were explored for further applications of face recognition.The major contributions of this dissertation are summarized below:(1)A hierarchical fully convolutional network(HFCN)based face detection method and a spatial transformer network(STN)based landmark detection method were proposed.In HFCN,hierarchical feature layers with different resolutions are used to detect different scale faces in complex backgpround scene,which avoid the construction of image pyramid.It can be trained and tested straightforward in an end-to-end manner and get an efficient performance.As for STN-based landmark detection method,it aims to improve the robustness of face region initialization.Firstly,transformation parameters are calculated using STN according to the initial face bounding box.Then,the normalized face is cropped according to parameters and is used for accurately landmark detection,which can alleviate the dependence of initial face bounding boxes.(2)A landmark perturbation-based data augmentation method was proposed.It is able to generate a huge number of misaligned face images,including stretched,distorted,rotated,cropped.The training dataset can be enriched and expanded by adding these artificial images.The unrobustness of the trained model to the above mentioned factors especially face recognition degradation problem caused by misalignment can be effectively alleviated.(3)A latent face model was proposed.First,it creates mappings from a hidden space to different media space.Then,the EM algorithm is adopted for solving the parameters.Finally,a coupled Joint Bayesian model is used to calculate the joint probability of two faces from different media.This method has a strong discriminative ability for learning intra-class and inter-class variations in each medium.Moreover,it can build associations with multi-source media and has a strong generalization ability.