Research On Cross-Age Face Recognition Based On Age-Invariant Identity Feature

With the development of the times,artificial intelligence has become one of the hottest topics in the contemporary era.Cross-age face recognition,as a branch of face recognition,has great practical significance and theoretical value in real life,such as finding children who have been lost for many years,identifying criminals who have absconded for many years,and passport identity verification in customs,etc.Although cross-age face recognition has achieved remarkable results in recent years,due to the complexity of face aging,cross-age face recognition is still facing great challenges.Face aging is a complex and continuous non-linear variation process.The appearance of the face will change significantly with aging in terms of shape and texture,and the intra-class distance may be even greater than the inter-class distance.Also,the aging process of different person is different,so modeling the face aging is a very complicated and difficult process.To solve this problem,this thesis has carried out related research.This thesis combines the direct sum of subspaces with a multi-task convolutional neural network that simultaneously performs face recognition and age classification tasks,and then proposes a model named Feature Subspace with Direct Sum multi-task Convolutional Neural Network(FSDS-CNN).By using multi-task learning,we can get both identity-related feature and age-related feature at the same time.This thesis then adds the direct sum module to these two features to conduct direct sum constraint for the two feature subspaces,so that the redundant component between the two features are effectively removed to a certain extent,that is,the age component within the identity feature is effectively culled.As a result,the correlation between identity feature and age feature is effectively reduced.In the end,we can obtain the age-invariant face identity feature that robust to the change of age.In addition,this thesis also makes a rigorously theoretical analysis for the introduction of the direct sum of subspaces into the field of cross-age face recognition,and introduces in detail how to implement the direct sum of subspaces in cross-age face recognition.We have conducted a large number of face recognition and face verification experiments on three public-domain benchmark aging datasets(Morph Album 2,CACD and Cross-Age LFW)after pre-training the model on the CASIA-WebFace dataset,and the two important hyperparameters of the direct sum module are explored.The experimental results are analyzed in detail and compared with some mainstream cross-age face recognition methods in recent years.The experimental results show that the FSDS-CNN model and its direct sum module proposed in this thesis can effectively reduce the correlation between identity feature and age feature,and thus can extract the age-invariant face identity feature that is robust to age change,so the performance of the cross-age face recognition model can be improved,which fully demonstrates the effectiveness and powerful generalization ability of our model.