Cross Age Face Verification Based On Deep Learning

Face recognition is widely used in daily life because it can recognize without sense.However,the facial features of the face change with age,which directly has a serious impact on the performance of face recognition,resulting in the phenomenon of unrecognized faces or incorrect recognition.How to suppress the age factor to extract age-changing facial features is the key to cross-age face recognition tasks.This paper aims at extracting age-invariant identity features,and introduces traditional algorithms and age estimation algorithms as auxiliary mechanisms to complete cross-age face verification tasks.The research content is mainly divided into the following parts:(1)A new data set was constructed to solve the problem of the limitations of the sample and the uneven age distribution in the FG-NET and MORPH Album2.The data in the FG-NET whose age is between 0 and 24 years old account for about79.24%.The MORPH Album2 has a small age span of the same person sample,which has certain limitations.Therefore,a new data set with an age span of 0-77 years old is obtained by taking out partial samples from the two data sets for fusion,thereby improving the robustness and generalization of the algorithm.(2)A cross-age face verification method that combines traditional algorithms and deep learning algorithms is proposed.First,the traditional algorithm LBP is improved,and the dual-coded average local binary pattern(DCALBP)algorithm is proposed.The texture feature and shape feature of the face are extracted using the DCALBP algorithm and the Histogram of Oriented Gradient(HOG)algorithm,and then the age feature vector is obtained by introducing canonical correlation Analysis(CCA)fusion of texture features and shape features.The Siamese Net is selected as the basic structure of the model,and the two network branches share weights.The CNN module in the Siamese Net is replaced with Res Net to extract facial features.Finally,the age feature is separated from the face features to obtain a set of age-invariant identity features,and then feature matching is performed to determine whether the input image pair is the same person.Experiments show that using DCALBP and HOG algorithms as auxiliary mechanisms to extract age features can effectively reduce the impact of age factors on the face verification system,and improve the robustness and accuracy of face verification.(3)A cross-age face verification method based on age estimation is proposed.This method introduces the idea of multi-task learning,realizes both the age estimation task and the cross-age face recognition task in a network,and uses the age features extracted from the age estimation task to assist the realization of the cross-age face verification task.The Siamese Net is selected as the basic structure of the model,and the two network branches share weights.The CNN module in the Siamese Net is replaced with Res Net,each branch network realizes age estimation and cross-age face verification at the same time.The two tasks share the low level characteristics.In the high level,each task learns its own characteristics in a targeted manner.In the branch network of the age estimation task,the extracted features are highly correlated with age,and the age feature vector is obtained.Then,the age feature is separated from the identity authentication task,and the age-invariant identity feature is obtained to match,so as to achieve cross-age face verification.Experiments on cross-age face data sets verify the effectiveness of the method.Using age estimation to assist cross-age face verification,the features extracted from the two tasks are mutually constrained,which improves the accuracy of face verification while also improving the generalization of the network model.