基于RDN和WGAN的图像超分辨率重建模型

Image super⁃resolution reconstruction model based on RDN and WGAN

文中针对现有基于生成对抗网络的单图超分辨率重建模型训练不稳定,以及重建后的图像细节视觉效果不理想等问题,提出基于残差密集网络(RDN)和WGAN的图像超分辨率重建模型。模型使用残差密集块(RDB)作为生成器的基本结构单元,通过残差密集连接的卷积层提取丰富的细节特征,以提高重建后的图像质量;引入WGAN的思想,以EarthMover距离定义对抗损失以解决SRGAN模型的训练不稳定问题;采用Charbonnier损失函数衡量重建图像与高分辨率图像的相似程度,以提高重建图像的视觉效果。通过在Set5,Set14数据集上的实验表明,相比于双三次插值、SRCNN、RDN、SRGAN等SR重建模型,该文模型重建后图像的PSNR和SSIM评价指标更高,图像亮度信息、纹理细节及视觉效果均有提升。

In order to solve the problems of unstable training of single⁃image super⁃resolution reconstruction models based on generative adversarial networks and unsatisfactory visual effects of reconstructed image details,an image super⁃resolution reconstruction model based on residual dense network(RDN)and Wasserstein GAN(WGAN)is proposed.In the model,the residual dense block(RDB)is used as the basic structural unit of the generator,and rich detailed features are extracted through convolutional layer with residual dense connection to improve the quality of the reconstructed image;the idea of WGAN is brought in,and the Earth⁃Mover distance is used to define the confrontation loss to solve the unstable training problem of the SRGAN model;the Charbonnier loss function is used to measure the similarity of the reconstructed image and the high⁃resolution image to improve the visual effect of the reconstructed image.The results of experiments on the Set5 and Set14 data sets show that,in comparison with Bicubic interpolation,SRCNN,RDN,SRGAN and other SR reconstruction models,the images reconstructed by this model have higher PSNR and SSIM evaluation indicators,the image brightness information,texture details and visual effects of this model have been improved.