基于视觉注意力机制的人脸属性迁移

人脸属性迁移作为计算机视觉领域的一个研究热点,对于数字娱乐制作、辅助人脸识别等领域有着重要的意义.现有的算法存在着生成图像模糊、转移属性无关区域变化等问题.针对这些不足,提出一种基于视觉注意力生成对抗网络的人脸属性迁移模型.生成器为减小属性无关区域的变化,引入视觉注意力分别输出RGB图像和注意力图像,并通过一定的融合方...

FACIAL ATTRIBUTES TRANSFER BASED ON VISUAL ATTENTION

As a research hotspot in the field of computer vision,facial attributes transfer is of great significance to digital entertainment production,auxiliary face recognition and other fields.In the existing algorithms,there are some unresolved problems such as generating image blur and transfer attribute-independent region change.To solve these problems,a facial attributes transfer model based on visual attention generative adversarial networks is proposed.In order to reduce the change of the attribute-independent region,the generator introduced the attention mechanism to output the RGB image and the attention image respectively,and obtained the attribute migration result through a certain fusion method.The multi-scale discriminator was used to maintain the details of high-dimensional feature mapping.We added cycle-consistent loss and attention image loss to the constraints to maintain the face identity information and focus on the migration of the relevant regions of attributes.The experimental results show that the proposed model can reduce the change of attribute irrelevant region and improve the effect of facial attributes transfer.