摘要:单幅图像的超分辨率重建（single image super-resolution，简称SR）是一项重要的图像合成任务.目前，在基于神经网络的SR任务中，常用的损失函数包括基于内容的重构损失和基于生成对抗网络（generative adversarial network，简称GAN）的对抗损失.但是，基于传统的GAN的超分辨率重建模型（SRGAN）在判别器接收高分辨率图像作为输入时，输出判别信号不稳定.为了缓解这个问题，在SRGAN以及常用的VGG重构损失框架上，设计了一个稳定的基于能量的辅助对抗损失，称为VGG能量损失.该能量损失使用重构损失中的VGG编码部分，针对VGG编码设计相应的解码器，构建一个U-Net自编码结构VGG-UAE，利用VGG-UAE的重构损失表示能量，并使用该能量函数为生成器提供梯度；基于追踪能量函数的思想，VGG-UAE使生成器生成的高分辨率样本追踪真实数据的能量流.实验部分验证了使用VGG能量损失将比使用传统的GAN损失可以生成更有效的高分辨率图像.



Abstract:Single image super-resolution (SR) is an important task in image synthesis. Based on neural nets, the loss function in the SR task commonly contains a content-based reconstruction loss and a generative adversarial network (GAN) based regularization loss. However, due to the instability of GAN training, the generated discriminative signal of a high-resolution image from the GAN loss is not stable in the SRGAN model. In order to alleviate this problem, based on the commonly used VGG reconstruction loss, this study designs a stable energy-based regularization loss, which is called VGG energy loss. The proposed VGG energy loss in this study uses the VGG encoder in the reconstruction loss as an encoder, and designs the corresponding decoder to build a VGG-U-Net auto encoder:VGG-UAE; by using the VGG-UAE as the energy function, which can provide gradients for the generator, the generated high-resolution samples track the energy flow of real data. Experiments verify that a generative model using the proposed VGG energy loss can generate more effective high-resolution images.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/6053