关键词: 人体姿态迁移, 图像生成, 生成对抗网络（GAN）, 深度神经网络

Abstract: Portrait synthesis guided by pose presents a challenging frontier in image generation. In the latest research, this paper has proposed an innovative network called the Step network, specifically designed to overcome the limitations identified in previous works. This approach differs from traditional methods by honing in on the spatial structure of the pose, enabling a gradual migration of the pose while minimizing the loss of spatial structure information at each step. Moreover, drawing inspiration from the triplet loss, a style discriminator is incorporated to enhance the quality of texture generation. In contrast to prior research, the paper has placed greater emphasis on refining the generation of facial areas. To achieve this, a specialized loss function is employed during the training process, combining triplet loss and L1 loss to optimize facial features, resulting in images that are more aligned with human perception. To evaluate the quality of the generated images, the paper utilizes evaluation metrics such as PSNR, SSIM, FID, and LPIPS. Through both qualitative and quantitative experiments comparing the approach with state-of-the-art models, the paper has demonstrated significant improvements across these metrics, confirming its superiority. Specifically, this method achieves a PSNR of 18.037 6, SSIM of 0.768 6, FID of 10.810 2, and LPIPS of 0.166 5.

Key words: human pose transfer, image generation, generative adversarial network(GAN), deep neural network