基于类别注意实例归一化机制的人脸年龄合成

近年来,生成对抗网络(generativeadversarialnetwork,GAN)家族已在人脸年龄合成任务上取得了巨大的成功.然而,通过研究发现,在解决人脸年龄合成的问题时,即使是善于利用年龄先验信息的条件生成对抗网络(conditional generative adversarial network, CGAN),重要的人脸年龄相关信息在一程度上也会被丢弃.这是导致以CGAN为代表的GAN家族在人脸年龄合成上的性能到达瓶颈期的一个重要因素.为此,提出了一种类别注意实例归一化机制(class-aware instance normalization, CAIN).该机制能够灵活地嵌入到CGAN中,形成一种新的生成对抗网络模型,即CAIN-GAN.CAIN-GAN能够充分利用人脸年龄先验信息来进一步提高人脸年龄合成性能.在公开数据集上的实验结果表明,与其他几种GAN家族的方法对比, CAIN-GAN方法仅通过利用人脸年龄相关信息就能对人脸年龄合成性能进行提升.

Class-aware Instance Normalization Mechanism for Face Age Synthesis

In recent years, the generation adversarial networks (GAN) family has been successfully applied for face age synthesis. Nevertheless, it is found that even if the conditional generation adversarial networks (CGAN) are good at using age prior information, the important age information will be discarded to some extent, when CGAN is used to address the problem of face age synthesis. This is an important factor that makes the performance of the GAN family represented by CGAN in face age synthesis task reach the bottleneck period. Therefore, a class-aware instance normalization (CAIN) is proposed, which can be flexibly embedded in CGANs, called CAIN-GAN, for thoroughly leveraging the age prior information to improve the performance of face age synthesis. Experiments on the public datasets show that the proposed CAIN-GAN can improve the performance of face age synthesis only by leveraging the face age-related information, compared with several GAN-based face age synthesis methods.