基于StarGAN的多属性风格图像生成的轻量化网络

Lightweight Network for Multi-Attribute Style Image Generation Based on StarGAN

生成对抗网络已广泛用于图像到图像的翻译任务,其中多属性变换得到了越来越多的研究和应用,目前网络架构的参数多而且模型复杂,需要较高的计算能力和存储成本;网络压缩技术如蒸馏和剪枝,主要侧重于视觉识别任务,很少实现对生成任务的压缩。本文提出了一种利用StarGAN的低级和高级特征训练参数较少的学生网络(stuStarGAN)的方法,首先采用知识蒸馏对生成器进行蒸馏,并设计学生判别器让教师判别器蒸馏学生判别器;然后在学生网络设计中采用skip-connection进行跨模块的特征融合;接着增加内容损失函数保持生成图像和原图像的内容信息的一致性;最后采用深度可分离卷积进一步降低参数量并提高图像生成质量。在CelebA和Fer2013数据集上的实验结果表明:模型能够在保证生成质量不降低的情况下,用较少参数生成多属性风格的图像,可以方便地移植到多种应用场景。

The generation countermeasure network has been widely used in image-to-image translation tasks,in which multi-attribute transformation has been studied and applied increasingly.However,the existing network architecture has many parameters and complex models,requiring high computing and storage costs;the traditional network compression technology mainly focuses on visual recognition tasks,and rarely implements the compression of generated tasks.Therefore,in this article we propose a method stuStarGAN to train the student network with fewer parameters by learning the low-level and high-level features of StarGAN.In our proposed method,first,we distill the generator with knowledge distillation,and design the student discriminator so that the teacher discriminator distills the student discriminator;then in the student network design,skip-connection is used to provide cross module feature fusion;second,the content loss function is added to keep the consistency of the content information between the generated image and the original image;finally,depth separable convolution is used to further reduce the number of parameters and improve the quality of image generation.The experimental results on benchmark datasets showed that the model could generate multi-attribute style images with fewer parameters without reducing the generation quality,making it easy to transplant to various application scenarios.