模式搜索星型生成对抗网络下的面部表情生成  

Facial expression generation under mode seeking StarGAN

作　　者：李萌萌 马力 贾宇峰 LI Mengmeng;MA Li;JIA Yufeng(School of Computer,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学计算机学院,陕西西安710121

出　　处：《传感器与微系统》2022年第7期147-151,共5页Transducer and Microsystem Technologies

摘　　要：针对生成对抗网络(GAN)在训练过程中因容易出现的模式崩塌现象导致面部表情生成图像效果不佳的问题,提出一种基于模式搜索星型生成对抗网络(StarGAN)的面部表情图像生成方法。首先将模式搜索正则项与星型生成对抗网络的生成器损失相结合,以改善星型生成对抗网络的模式崩塌现象;其次将网络结构中生成器的普通卷积层使用空间可分离卷积替代,相对减少训练参数,提高模型训练的稳定性。实验在2个数据集上与CycleGAN,DiscoGAN和StarGAN等模型生成的图像进行质量比较,在CK+数据集上弗雷歇距离分别提高了2.98,2.20和1.64;在FER2013数据集上弗雷歇距离分别提高了2.60,1.72和0.68。本文所提出的模式搜索StarGAN模型可以更有效改善模式崩塌问题,进而提高表情图像的质量和模型训练的稳定性。Aiming at the problem of poor facial expression generation images caused by the collapse of the pattern during the training process of the generation adversarial network(GAN),propose a facial expression image generation method based on mode seeking StarGAN(MS-StarGAN).Firstly,the network combines the mode seeking regular term with the generator loss of the StarGAN to improve the mode collapse of StarGAN.Secondly,the common convolutional layer of the generator in the network structure is replaced by a spatially separable convolution,which relatively reduces the training parameters and improves the stability of the model training.The experiment is compared with the latest 4 methods on 2 datasets.In the CK+dataset,the Fréchet Inception Distance is improved by 2.98,2.20,and 1.64,compared to CycleGAN,DiscoGAN,and StarGAN,in the FER2013 dataset,compared with CycleGAN,DiscoGAN and StarGAN,the Fréchet Inception Distance is improved by 2.60,1.72 and 0.68.The proposed mode seeking StarGAN model can more effectively improve the mode collapse problem,and then improve the quality of expression image and the stability of model training.

关 键 词：星型生成对抗网络 模式崩塌 搜索正则项 图像生成 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]