Generation Model of Character Posture TransferBased on Self-attention Mechanism

This paper proposes a character pose transfer generative model fused with the self-attention mechanism to address the issues of loss of texture details and unreasonable pose transfer in images generated by character pose transfer. First, based on the two-stage pose transfer generative model, the improved self-attention module is introduced into the generative adversarial network to reduce the interaction between similar features, which improves the ability to learn texture details and capture information, enhances saliency modeling of posture features. Then, the Markov discriminative model is used to enhance the ability to discriminate the details of the generated image. Finally, the optimized content loss function is used to constrain the image feature information loss of the entire model, promote semantic consistency between the generated and the real images, and strengthen the rationality of pose transfer. The experimental results demonstrate that, compared with the PG2 method on the DeepFashion and Market-1501 datasets, the IS and SSIM values of our model has increased in 0. 388 and 0. 032, 0. 036 and 0. 065, respectively.