Research On Image Synthesis Methods Of Human Body Parts Based On Generative Adversarial Network

In the fields of medical rehabilitation,human-computer interaction,and public security,it is of great significance to use computer technology to understand images of human body parts.For this reason,some facial attribute synthesis and editing methods and 3D hand pose estimation methods have been proposed,and are gradually widely used in related fields.Although the existing facial attribute synthesis and editing methods can effectively synthesize the target face image according to the specified attribute labels,the quality of the synthesized target face image still needs to be improved.To solve this problem,we propose a novel multi-task method called ARU-GAN for facial attribute synthesis and editing.In addition,in order to further study the human hand image,we improve the ARU-GAN method and propose a 3D hand pose estimation and hand depth image synthesis method called HDR-GAN.The main contributions are as follows:1.To solve the problem that the current facial attribute synthesis and editing methods are difficult to synthesize high-quality target face images according to specified attribute labels,we propose a multi-task method called ARU-GAN based on ARU-net.The ARU-net in this method introduces a skip connection based on the encoder-decoder architecture to effectively fuse different level features of the input face image,and uses an adversarial regularization term to constrain latent variables during training,thus synthesizing high-quality target face images in facial attribute synthesis and editing tasks.To further improve the quality of the synthesized target face image,ARU-GAN uses the target face image synthesized by the above tasks to jointly train the discriminator.The experimental results show that ARU-GAN can synthesize high-quality target face images according to the specified attribute labels in the facial attribute synthesis and editing tasks.2.By improving the above method,we propose a 3D hand pose estimation and hand depth image synthesis method HDR-GAN,consists of the 3D hand pose estimation network HPE-GAN and the hand depth image synthesis network HDG-GAN.In HDR-GAN,the HPE-GAN composed of an encoder and a discriminator performs more accurate 3D hand pose estimation on the occluded hand in the RGB image through adversarial training.At the same time,HDG-GAN composed of the decoder and the discriminator can synthesize a given 3D hand pose into a corresponding hand depth image through the same adversarial training pattern.Finally,we qualitatively and quantitatively analyze the results of 3D hand pose estimation and hand depth image synthesis,and use HDR-GAN to initially implement the task of reconstructing hand depth images based on RGB images.