Cost-sensitive Face Recognition Based On Residual Neural Network

Face recognition is one of the classic tasks in the field of image processing and computer vision.Its purpose is to enable computers to extract facial features through specific algorithms and programs to identify the identity of the target face.Most of the existing face recognition algorithms aim to reduce the misclassification rate,and the inter-class error cost is not taken into consideration.However,in practical applications,different kinds of misclassifications are often not equivalent.Therefore,it is necessary to introduce cost-sensitive into face recognition.Based on the existing residual neural network,this paper proposes an improved residual neural network structure which enhances the feature expression power of the model.It also introduces a cost-sensitive cross entropy loss function which takes both accuracy and cost in consideration and can be used to solve multi-class cost-sensitive problems.Firstly,the face feature extraction network is studied.by comparing accuracy,memory and speed,the residual neural network is chosen to be the best face feature extraction network.Then,multi-scale feature fusion and channel-wise attention are integrated into the residual neural network,which further enhances he feature expression ability of the network.Related experiments are conducted to prove its effectiveness.Finally,a cost-sensitive cross entropy loss function is proposed,which significantly reduces the misclassification cost almost without any loss of precision.The effectiveness of the improved loss function is verified by designed experiments.By introducing multi-scale feature fusion and channel-wise attention into the residual neural network,the feature expression ability of the network model is enhanced.By introducing the cost-sensitive mechanism into the cross-entropy loss function,the misrecognition cost can be reduced under the premise of maintaining accuracy,which can be used to solve multi-class cost-sensitive problems and makes a contribution to the cost-sensitive face recognition field.