基于GANs和迁移学习的人脸表情生成方法研究  

Facial Expression Generation Based on GANs and Transfer Learning

作　　者：马志萍 Ma Zhiping(School of Information,North China University of Technology,Beijing,100144,China)

机构地区：[1]北方工业大学信息学院,北京100041

出　　处：《仪器仪表用户》2021年第2期15-18,共4页Instrumentation

摘　　要：生成对抗网络(GANs)自2014年被提出以来,被广泛应用在文本、图像等领域,并对其进行生成方面的研究。但是最近的研究发现,生成对抗网络在有限样本的情况下无法取得好的生成效果,产生过拟合等问题。为了解决这个问题,本文提出了一种利用迁移学习和改变网络结构相结合的方法。与传统的迁移学习不同的是,在对源网络卷积层进行迁移过程的同时,引入AdaFM模块并用Style Blocks替换生成器前两层,使迁移后参数更好地适应目标网络,从而提升在小样本的情况下的生成效果并降低结构复杂度,加快训练速度。本文分别使用CelebA人脸数据库和RaFD人脸表情数据库作为源任务和目标任务的训练数据对该算法进行测试。测试结果表明,所采用的方法在有限样本的条件下,能够使图片具有更好的生成效果。Since GANs was proposed in 2014,it has been widely used in the generative research of text,images and other fields.However,recent work has shown GANs can’t generate highly images on limited data,leading to overfitting and other problems.To alleviate this problem,we propose to leverage the method of combining transfer learning and network structure.Different from the traditional transfer learning,we introduce AdaFM module and the first two layers of the generator are replaced with Style Blocks,aiming at boosting performance on the limited data and simplied structural complexity in order to speed up training while transferring filters from source network.We evaluate the algorithm with CelebA face database and RaFD facial expression database as source task and target task,respectively.The experiment results show that GANs can generate highly images using limited-data by this way.

关 键 词：表情生成 深度学习 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]