融合衣物特征的人体姿态迁移方法

Design of human posture migration network based on clothing features classification

人体姿态迁移任务即将输入图片中的人物姿势转变为目标姿势,同时保证图片中衣物等外貌特征不变。本文提出了一种新型的生成式对抗网络用来生成人体姿态迁移后的图片,对抗网络的生成器使用注意力机制和风格迁移方法交替对人的外貌信息和姿势信息进行更新。此外,本文较以往的研究还增加了人物衣物的平铺图片作为网络输入,以此来补充人体姿态迁移过程中可能丢失的外貌信息,网络能够使生成图片中人物所穿衣物更加接近真实样式。通过大量实验证明,本文提出的网络在量化分数上获得了更高得分,图片的肉眼观感也更真实自然。

The human pose transfer task is to change the human body in the input picture into the target posture, while ensuring that the appearance features such as clothes remain unchanged. In this paper, a new generative adversarial network is proposed to generate pictures of human posture after migration. The network generator uses attention mechanism and Style transfer block to update the appearance information and posture information alternately. In addition, compared with the previous research, this paper also adds the tile pictures of people′s clothes as the network input to supplement the appearance information that may be lost during the migration of human posture. The network can more realistically and naturally restore the clothing style worn by people. Through a large number of experiments, it is proved that the network proposed in this paper shows a higher score in the quantitative score, and the visual perception of the picture is more realistic and natural.