Transformer-based Continuous Face Age Editing And Generation Method

With the continuous development of modern multimedia information technology,the popularity of face image shooting and acquisition,editing,analysis,and recognition has been prevalent.As a representative global feature of the face,the facial image of each person at different ages carries rich and different information.Face age editing techniques,an integral branch of face attribute editing,commonly employ generative adversarial networks(GANs)or Transformers for age modification and prediction in face generation.This technology holds significant potential and practical value,finding applications in intelligent film and television,entertainment,social networking,public security criminal investigation,and big data domains.Although deep neural networks have experienced rapid development,both traditional methods and deep network approaches for face age editing encounter the challenge of achieving natural editing transitions and maintaining continuity and authenticity in continuous age face image editing.To address these challenges,this thesis focuses on addressing the editing and generation challenges in continuous age face images using Transformer,leveraging its model characteristics.The study highlights the following key features:(1)A Transformer-based feature mapping model(TPN)is proposed to address the challenge of extracting and fusing detailed features from face images across various age groups.The model enhances the representation of latent space and hidden vectors for capturing global facial age features,effectively guiding the face editing process.Experimental results demonstrate that the proposed method partially mitigates the limitations of convolutional operations in capturing global features of face images.Additionally,it lays a solid foundation for achieving high-quality editing and generation of continuous age face images.(2)In response to the challenge of preserving attribute features and details during age editing,this thesis presents Former Age,a progressive bi-directional age editing generation model.Former Age leverages the capabilities of the Transformer architecture and introduces a novel approach to integrating detailed information from the potential space into the progressive generator.Comparative experimental results demonstrate that the model is more effective in terms of identity preservation,age editing,and image quality.(3)Former Age++ is proposed to address the current problems of age discontinuity and poor stability of cross-ethnic editing in face age editing tasks.The enhanced version builds upon the foundations of Former Age and conducts extensive optimization,carries out research on finer-grained continuous age face image generation,starts from both data input and output,further adjusts the loss function,training strategy,improves the model details around a specific face age global attribute editing task,and carries out a large number of ablation and comparison experiments to verify the effectiveness of the improvements,and finally also systematically reviews and analytically evaluates the model design process of this thesis.