Deep Convolutional Neural Network For Age-invariant Face Recognition

Face recognition technology is one of the important and widely used research topics in the field of computer vision and biometric recognition,which has been developed for decades.In recent years,researches on face recognition have focused on the problem of interference of illumination,expression,posture and age.Among them the field of age invariant face recognition which is face recognition based on face images with different age spans for the same person still faces many challenges.Nowadays,the research of age invariant face recognition has becoming increasing important.Moreover,it has wide application value,such as.searching for missing children,identifying absconding criminals,etc.In this paper,aiming at the problem of the change of facial features caused by age change the age-invariant face recognition algorithm is researched based on the basic principle of deep convolutional neural network.The research mainly includes the following aspects.First,face detection and preprocessing algorithms are studied.Aimed at the problem of the low quality of collecting face images caused by illumination,posture and equipment,the face images have been preprocessed by face landmarks detection,alignment and face preprocessing operation,which is benefit for later feature extraction.Second,in order to accelerate the convergence speed of deep convolutional neural network,the VGG16 network model has been improved by increasing the number of convolutional layers and introducing local residuals.Based on the training strategy of small data small network big data big network and the loss function of A-softmax,the network model was fitted.The deep and abstract information of face image can be extracted by this improved VGG16 network model.In order to further realize age-invariant face recognition,this paper proposes the algorithm that combines the VGG16 network model and face feature linear model.This algorithm is used to extract the relatively stable features of face image with the change of age,namely the identity featuresIn order to obtain the parameters of the linear model,this paper proposes the improved EM maximization iteration algorithm,which can increase the convergence speed of the EM algorithm.To obtain the high probability of face recognition,the face second match algorithm is studied in this paper.Finally,a large number of age-invariant face recognition experiments and comparative experiments are carried out on the public age FG-NET and MORPH-II database.Then,evaluation indexes are used to measure the performance of the algorithms.The experiment results verify the effectiveness of the proposed method.