Research On Face Recognition Method Under Unconstrained Condition

Face recognition technology with constraints has been applied in some specific business scenarios,while security and other important fields pay more attention to the recognition effect in real unconstrained scenes,so face recognition without constraints is the most urgent problem to be solved in this field.Therefore,this thesis studies the face recognition algorithm from the aspects of training data,loss function and network structure as follows:1.This thesis studies a method of automatically cleaning large-scale face training sets.Firstly,the interference samples that do not contain faces in the MS-CELEB-1M are cleared by the face detection algorithm in the preprocessing stage.Then,combined with the self-training idea,the network is trained on the data set,and after each training,the samples of each category are judged once.Finally,the sample is divided into two parts,the correct and the wrong,and then the cosine distance between each error sample and the center of the correct sample class is calculated,and the sample above the given threshold is recalled to complete the final cleaning.This thesis studies a loss function based on angular distance metric learning,which aims to eliminate the influence of sample imbalance and increase the discriminability of facial features.First,the author studies the essential causes of the sample imbalance,which can be mitigated by weight normalization but does not eliminate this effect.Therefore,this thesis reconstructs the Softmax cross entropy loss function in the form of triples,so that the scarce category only acts as a negative sample,and does not maintain its class center in the back propagation phase,thus avoiding the sample while using the scarce category information.Unbalanced impact.Finally,the feature is changed from the European space to the cosine space,so as to ensure the same standard as the test phase,and the angular separation is added to improve the discriminability of the feature.This thesis studies a kind of face recognition network structure based on attention supervision.The core idea is to use the attention mechanism to make the network pay more attention to the face attribute characteristics of the face.First of all,combined with the residual idea,the attention branch is introduced on the basis of the existing network,and the attention weight is obtained by using the original layer,so that no additional layer and a large number of parameters are needed.Then the attention feature is merged with the original feature to avoid the influence of the Sigmoid activation function and ensure the smooth end-to-end training.Finally,the attitude network is used to calculate the deflection angle of the sample,and the weight of the feature to be fused is adaptively adjusted according to the angle value,thereby further enhancing the ability of the attention mechanism to characterize the robustness of the attitude.This thesis verifies the widely used face recognition test standards of LFW,MegaFace and CFP.The excellent performance proves the effectiveness of the proposed algorithm.