Face Aging With Gradient Guide Generative Adversarial Network

In the field of image generation,face aging is an important research direction,and the task has a very high research value in the field of cross-age recognition research as well as current social general entertainment applications.However,the mission still has various difficulties at present.These difficulties include the lack of the same person over a long age range during the construction of the dataset,deep learning networks for this task are difficult to train and the difficulty of preserving the identity information of face images.The main target of this paper is to preserve the identity information while ensuring that the cross-age face generation task is completed intact.Therefore,in this paper,feature-guided generative adversarial networks based on VAE and conditional generative adversarial networks are constructed.In the past,the tasks for face images were mainly focused on face recognition and age prediction,and the tasks for face age classification were very sparse.This means that the training of age classifier faces difficulties such as sparse dataset and low annotation of dataset.In the paper,addressing the above problems by first investigating the past work.Then select a backbone network for face age classification.Subsequently,using the transfer learning method,the face recognition network is first trained using this network and based on this network,migration training is performed to transfer the original face recognition task to the age classification task.Then,a new strongly supervised approach,i.e.,feature-guided approach,is proposed to ensure that the generative model can better retain the identity information of face images.Meanwhile,the discriminator in the original generative adversarial network is split into two parts.One part is the traditional discriminator network,and the other part is the face age classification network,and the loss function used for network training is reconstructed.In the experimental process,age classification networks with different classification accuracies are used as part of the discriminator to train the generator,and it is found that higher or lower classification accuracies have a greater impact on the final results of the generator.Finally,the results of the images generated by this paper are compared with those of state-of-the-art algorithms,and the differences between the results of the proposed algorithm and those of past algorithms are analyzed.At last,a cross-age face image generation system is constructed based on the above model.