PDA-GAN:融合通道位置注意力机制和并行空洞卷积的人脸年龄合成网络
张珂1, 于婷婷1, 石超君2, 娄文硕2, 刘阳2(1.华北电力大学电子与通信工程系;2.华北电力大学) 摘 要
目的 人脸年龄合成旨在合成指定年龄人脸图像的同时保持高可信度的人脸,是计算机视觉领域的热门研究方向之一。然而目前主流人脸年龄合成模型过于关注纹理信息忽视了与人脸相关的多尺度特征,此外网络存在对身份信息筛选不佳的问题。针对以上问题,本文提出了一种融合通道位置注意力机制和并行空洞卷积的人脸年龄合成网络(PDA-GAN)。方法 PDA-GAN 基于生成对抗网络提出了并行三通道空洞卷积残差块和通道-位置注意力机制。并行三通道空洞卷积残差块将三种膨胀系数空洞卷积提取的不同尺度人脸特征融合,提升了特征尺度上的多样性和总量上的丰富度;通道-位置注意力机制通过对人脸特征的长度、宽度、深度显著性计算,定位图像中与年龄高度相关的通道和空间位置区域,增强了网络对通道和空间位置上敏感特征的表达能力,解决了特征冗余问题。结果 实验在 FFHQ 数据集上训练,在 Celeba-HQ 数据集上测试,将本文提出的 PDA-GAN 与最新的3种人脸年龄图像合成网络进行定性和定量比较,以验证本文方法的有效性。实验结果表明 PDA-GAN 显著提升了人脸年龄合成的身份置信度和年龄估计准确度,具有良好的身份信息保留和年龄操控能力。结论 本文方法能够合成具有较高真实度和准确性的目标年龄人脸图像。
关键词
PDA-GAN:Face age synthesis network fusing channel location attention mechanism and parallel dilated convolution
(NCEPU) Abstract
Objective Face age synthesis aims to synthesize face images of specified age while maintaining high-confidence faces, which is one of the hot research directions in the field of computer vision. However, the current mainstream face age synthesis models focus too much on texture information and ignore the multi-scale features related to faces. In addition, the network has the problem of poor screening of identity information. In response to the above problems, this paper proposes a face age synthesis network (PDA-GAN) that fuses the channel and spatial position attention mechanism and parallel dilated convolution. Methods PDA-GAN proposes a parallel three-channel dilated convolutional residual block and a channel-position attention mechanism based on generative adversarial networks. The parallel three-channel hole convolution residual block fuses the different scales of face features extracted by the hole convolution with three expansion coefficients, which improves the diversity of feature scales and the total richness; the channel-position attention mechanism The length, width, and depth saliency calculation of facial features locates the channels and spatial location regions that are highly correlated with age in the image, enhances the network""s ability to express sensitive features on channels and spatial locations, and solves the problem of feature redundancy. Results The experiments were trained on the FFHQ dataset and tested on the Celeba-HQ dataset. PDA-GAN proposed in this paper was qualitatively and quantitatively compared with the latest three face age image synthesis methods to verify the effectiveness of this method. The experimental results show that PDA-GAN significantly improves the identity confidence and age estimation accuracy of face age synthesis, and has good identity information retention and age manipulation ability. Conclusion The method in this paper can synthesize face images of target age with high realism and accuracy.
Keywords
|