摘要： 跨年龄人脸合成是指通过已知特定年龄的人脸图像合成其他年龄段的人脸图像，在动漫娱乐、公共安全、刑事侦查等领域有广泛的应用。针对跨年龄人脸合成图像容易产生器官变形扭曲、人脸局部特征保持效果不佳等问题，提出一种基于条件对抗自动编码器的合成方法。通过在解码器结构中引入通道关注和空间关注模块，分别从通道域和空间域提取重要信息，使模型在训练过程中忽略背景等无关信息，聚焦人脸图像变化的区域，有效解决合成图像器官扭曲变形等问题。此外，设计一种多尺度特征损失网络，从多个尺度更深层次地约束人脸图像的局部结构特征，从而保持人脸合成过程中局部特征结构的稳定性。在UTKFace跨年龄人脸数据集上的实验结果表明，与CAAE方法相比，该方法有效避免了人脸器官变形扭曲问题，能够更好地保持人脸局部结构特征，具有较佳的人脸合成效果和细节保持能力。

关键词: 跨年龄人脸合成, 条件对抗自动编码器, 通道关注模块, 空间关注模块, 多尺度特征损失网络

Abstract: Cross-age face synthesis involves synthesizing facial images of other age groups from facial images of known specific age groups.It has a wide range of applications in the fields of animation entertainment, public safety, criminal investigation, and so on.To solve the problems of organ distortion and poor local feature preservation in cross-age face image synthesis, a cross-age face image synthesis method based on a Conditional Adversarial AutoEncoder (CAAE) is proposed.By introducing both channel and spatial attention into the decoder structure, more important parts are taken from the channel and spatial domains, respectively, so that the model ignores irrelevant information such as the background in the training process, focuses on the changing area of the face image, and effectively avoids the distortion and deformation of organs in synthetic images.In addition, a multi-scale feature loss network is designed to constrain the local structural features of face images from multiple scales to maintain the stability of the local feature structure in the face synthesis process.The experimental results from the UTKFace cross-age face dataset show that compared with the CAAE method, this approach effectively prevents the deformation and distortion of facial organs, can better maintain the local structural features of the face, and has a better face synthesis effect and detail retention ability.

Key words: cross-age face synthesis, Conditional Adversarial AutoEncoder(CAAE), channel attention module, spatial attention module, multi-scale feature loss network

中图分类号: 

• TP391