基于无监督学习的单人多姿态图像生成方法

针对目前视觉监控领域中采集到的人物数据样本量少和特征单一的问题,提出了一种具有高视觉感知约束的双向生成对抗网络生成期望人物姿态图像的方法。采用给定人物的单个图像和期望姿态的二维骨架作为双向生成对抗网络的输入,生成具有该目标人物期望姿态的图像。将生成的期望姿态图像反映射回原始姿态图像,利用少量的图像以无监督学习方式进行学习,生成该人物期望姿态的高质量图像。提出的方法在DeepFashion公开数据集上进行了实验,结果表明,采用文中提出的方法生成的图像结构相似度(SSIM)比以往的方法提高了0.28,有效的提升了基于无监督学习的单人多姿态人物图像生成的质量。

Single-personmulti-pose Image Generation Method Based on Unsupervised Learning

Aiming at the problems of small sample data and single feature of the person data collected in the field of visual surveillance, a method of generating a desired person pose image with a bidirectional generative adversarial network with high visual perception constraints is proposed. A single image of a given character and a twodimensional skeleton of a desired pose are used as inputs to a bidirectional generation adversarial network to generate an image with the desired pose of the target person. The generated expected pose image is mapped back to the original pose image, and a small number of images are used for learning in an unsupervised learning manner to generate a high-quality image of the character′s desired pose. The proposed method is tested on the DeepFashion public data set. The results show that the image structure similarity(SSIM) generated by the method is 0.28 higher than that of previous methods, which effectively improves the image generation quality of single-personmulti-pose based on unsupervised learning.