摘要
为解决基于深度卷积生成对抗网络的语义图像修复模型存在的重建结果内容、风格、细节特征还原不准确问题以及模型训练不稳定问题,提出一种结合残差块和注意力块的Multi-Loss GAN模型.同时,在图像生成阶段,向模型引入谱归一化和Wasserstein距离以稳定模型训练;在图像修复阶段,向模型增设差异网络和Vgg19特征提取网络分别提供差异、内容、风格损失协助模型寻找最优编码以提升最终的修复效果.最后在CelebA数据集上进行大量仿真实验.结果显示,Multi-Loss GAN较于DCGAN方法和GLCIC方法在PSNR和SSIM上分别提升0.6~2.0 db,0.01~0.05.
To solve the problem of inaccurate reconstruction of content,style,and detail feature of the semantic image inpainting model based on the deep convolutional generative adversarial network and the problem of instability of model training,a Multi-Loss GAN model combining residual blocks and attention blocks is proposed.Specifically,in the image generation phase,the spectral normalization and the Wasserstein distance are apply in the model to stabilize the model training,additionally,in the image inpainting phase,to help the model improve the final inpainting effect,a difference network which provides difference loss and a Vgg19 feature extraction network which provides content and style loss are added to seek optimal code.At last,a lot of simulation experiments have been carried out on the CelebA dataset.The results show that the Multi-Loss GAN is improved by 0.6~2.0 db and 0.01~0.05 respectively compared with the DCGAN method and the GLCIC method on PSNR and SSIM.
作者
曹真
杨云
齐勇
李程辉
CAO Zhen;YANG Yun;QI Yong;LI Cheng-hui(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an 710021,China)
出处
《陕西科技大学学报》
CAS
2020年第3期158-165,共8页
Journal of Shaanxi University of Science & Technology
基金
国家自然科学基金项目(61601271)
陕西省科技厅重点研发计划项目(2017NY-124)。