Font Size: a A A

Research On Human Face Age Transformation Method Based On Generative Adversarial Network

Posted on:2022-04-22Degree:MasterType:Thesis
Country:ChinaCandidate:W J YangFull Text:PDF
GTID:2518306569494684Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
With the development of the Internet and mobile devices,it has become popular to share personal images on social media,which has promoted the rapid development of face image editing technology.Traditional facial image processing technology is able to adding facial decorations and beautifying the faces.While the face image editing method based on adversarial generative networks can change the inherent attributes of face,such as age and gender.The age attribute editing technology of face images plays an important role in the fields of entertainment,social media,film and television media,and security.It can be widely used in some tasks,such as face image special effects,finding the missing children and synthesizing face recognition data across ages.Currently,face age transformation is still a challenging task.Traditional face aging simulation methods are generally based on average face interpolation and biological modeling.These methods suffer the problems of difficulty in data collection,low image quality,and high computational cost.Although the generative models have made great improvements in image quality and resolution,however,there are some limitations in the existing methods.For example,the texture of the synthesized human face is smooth,which is still far from the complex texture in reality.And they can't be able to transform skull shapes.In order to solve problems above,the work of this article is summaried as follows:Aiming at the two-way age transformation task,we analyze the GAN-based image translation methods,and propose a novel two-way age transformation model.The autoencoder-based generator is able to model the transformation of skin texture,facial soft tissue and skull.Using the disentangled representations,the model is able to retain identity information and realize unsupervised two-way face age editing without paired data.Aiming at the task of continuous age transformation,we proprose a method based on multi-task age feature enocder.By increasing the output branches of the age encoder and the discriminator,we build a single model for multi-task continuous age transformation.The content encoder encodes the face images into latent codes shared by multi-domains.The decoder adopts multi-scale features to synthesize the target image.Multi-scale feature fusion can control the overall image content and local texture details.The proposed method can improve the quality of the composite image,realizing the disentangled representation of attributes,and allow interpolation and editing operations in the latent space,so as to achieve fine-grained continuous age editing effects.The experiment results show that the proposed method achieves better image quality than the exists methods.Specifically,compared with the state-of-the-art methods,the proposed method sets a new state-of-the-art on several datasets.
Keywords/Search Tags:generative adversarial networks, face age transformation, face attribute editing, image translation
PDF Full Text Request
Related items