人脸属性编辑的全局组织网络算法

提出一种新的基于生成对抗网络的人脸属性编辑全局组织网络算法.人脸属性编辑是指通过结合编码解码器结构与生成对抗网络，生成具有期望属性的人脸图像.传统的编码解码器结构对人脸的重构和编辑能力有限.直接将编码特征与属性标签结合会因为融入编码特征造成属性编辑性能低下，同时，也会由于缺失编码特征造成人脸还原度的损失，两者无法平衡.因此，提出U型传递方式与全局组织单元. U型传递改变了传统的属性流动方式，并生成反向状态.全局组织单元结合反向状态生成全局属性信息，在编码解码器中搭建桥梁，帮助解码器更好地融入编码器特征与属性信息.与此同时，为了更好地配合全局组织模块，重新设计了编码器下采样.实验结果表明，本文所提方法可以同时提高模型的人脸重塑与属性编辑能力. 

Global Organization Network for Facial Attribute Editing

In this paper,a novel global organization network for facial attribute editing was proposed based on a generative adversarial network. Facial attribute editing is to generate face images with desired attributes by combining the encoder-decoder structure with GAN. However,the traditional encoder-decoder structure has limited ability to reconstruct face images and edit attributes. Directly combining the encoder features with the attribute label,the method can result in poor attribute editing performance due to the incorporation of the encoder features,while,face restoration degree degrades because of the absence of the encoder features,and the two can not be balanced. Therefore,the global organization units (GOU) and U-shaped transferring method were proposed. U-shaped transferring method was arranged to change the traditional attribute flow mode and generate inverted states. Combining with the inverted states,the global organization unit was used to generate global state,build a bridge between the encoder and decoder,and help the decoder better integrate encoder features and attribute information. Meanwhile,in order to better fit the global organization unit,an encoder down-sampling was redesigned. Experimental results show that the proposed method can improve the ability of face reconstruction and attribute editing simultaneously.