| Face 2D image generation is a hot research topic in the field of computer vision,and it has been widely applied in areas such as film and entertainment,AI security,virtual reality,and image restoration.The goal is to generate high-quality and realistic face images using image processing methods.The development of generative adversarial networks has propelled research in this field,achieving breakthroughs in image quality,realism,detail preservation,and attribute control.Face image generation involves multiple sub-tasks,and this thesis focuses on the research of facial expression synthesis and face super-resolution reconstruction.The main contributions of this work are as follows:To address the issue of detail distortion that may arise when using facial geometric features for expression generation,such as irregular skin texture,this thesis proposes a facial expression generation model that learns style and identity information.Given the potential differences in position,scale,and shape between the input facial image and the input facial contour,a facial contour alignment module is proposed to adjust and align the input facial contour based on the input facial image.This helps the model to match the local structure and details of the face.The aligned facial contour map and the input facial image are then fed into the facial generation module of the model.As other facial information is difficult to precisely describe with discrete labels like expressions,this thesis introduces facial style and identity discriminators.These discriminators extract style information(e.g.,skin tone,gender,and hairstyle)and identity information from the image pairs to facilitate feature discrimination,assisting the model in generating facial images that are consistent with the style and identity of the input facial image.We conducted facial expression generation experiments and ablation experiments on the Face Forensics++ and Ra FD datasets.We compared our model with the state-of-the-art methods on the Ra FD dataset.The results show that the proposed facial expression generation model in this thesis generates facial images with good realism and detailed information.The facial expression generation model proposed in this thesis reduces the issue of detail distortion.However,the generated facial images do not achieve high resolution and exhibit blurriness when enlarged.To further enhance the details and clarity of facial images,this thesis investigates face super-resolution reconstruction.Addressing the problem of accurately restoring facial image details without relying on prior facial information,this thesis proposes a cascaded upsampling face reconstruction model that incorporates multi-scale facial global attention.To improve the resolution and realism of the images,a cascaded upsampling network structure is introduced to progressively increase the resolution of the input facial image,gradually adapting to the target task,thereby enhancing the model’s robustness and generalization ability.Since facial prior information is not used,it is necessary to fully utilize the information between highresolution and low-resolution images.In each upsampling generative adversarial network,this thesis presents a multi-scale facial global attention mechanism.This mechanism calculates the similarity between each position in the extracted multi-scale feature maps,better utilizing contextual information and enhancing the ability to capture facial details.We trained the face super-resolution model on the FFHQ dataset,and tested the reconstruction effect on the images generated by the proposed expression generation model in this thesis.We further conducted ablation experiments on the Face Forensics++ dataset,and compared our face super-resolution model with the stateof-the-art face super-resolution methods on the Celeb A-HQ dataset.The experiments show that the proposed face super-resolution model can generate clearer and more detailed images without relying on additional facial prior information. |
