双判别器生成对抗网络图像的超分辨率重建方法

Image Super-Resolution Reconstruction Method Using Dual Discriminator Based on Generative Adversarial Networks

提出一种可用于改进图像超分辨率重建质量的双判别器超分辨率重建网络(DDSRRN)。该网络在生成式对抗网络(GAN)的基础上增加一个判别器,将Kullback-Leibler(KL)和反向KL散度组合成一个统一的目标函数来训练两个判别器,利用这两种散度的互补统计特性,能在多模式下分散预估计密度,从而避免重建过程中网络模型的崩溃问题,提高模型训练的稳定性。针对模型损失函数的设计部分,首先使用Charbonnier损失函数来构建内容损失,利用网络中间层的特征信息来设计感知损失和风格损失,最后为缩减图像重建时间,在网络结构中引入反卷积来完成图像重建操作。实验结果表明本文方法在主观视觉上具有丰富的细节,获得了更好的主观视觉评价和客观量化评价,网络泛化能力好。

In this study,we propose a dual discriminator super-resolution reconstruction network(DDSRRN)that can improve the super-resolution reconstruction quality of images.By adding a discriminator based on generative adversarial networks,the DDSRRN combines the Kullback-Leibler(KL)divergence and reverse KL divergence into a unified objective function for training two discriminators.Thus,the complementary statistical properties obtained from these divergences can be exploited to effectively diversify the pre-estimated density under multiple modes.Additionally,model collapse is effectively avoided during the reconstruction process,and the model training stability is improved.The model loss function can be designed based on the Charbonnier loss function to estimate the content loss.Furthermore,the intermediate features of the network are used to design the perceptual loss and style loss.Finally,a deconvolution layer is designed to reconstruct the super-resolution images,thereby reducing the image reconstruction time.The proposed method is experimentally demonstrated to provide abundant details.Thus,the proposed method exhibits good generalization ability and obtains improved subjective visual evaluation and objective quantitative evaluation.