Edge-Informed Single Image Super-Resolution

• Edge-Informed Single Image Super-Resolution

  • 性能：
  • 把sr任务描述为一个 in-between pixels inpainting task，并分为两个阶段，一个是edge enhancement，一个是image completion，两个阶段用两个GAN来完成，两个G都是encoder-decoder，先downsample两次再upsample，其它一些细节如下

  We use dilated convolutions in our residual layers. Our generators follow similar architectures to the method proposed by Johnson et al. [21] shown to achieve superior results for super-resolution [35, 14], image-to-image translation [45], and style transfer. Our discriminator follows the architecture of a 70 × 70 PatchGAN [20, 45] that classifies overlapping 70×70 image patches as real or fake. We use instance normalization [40] across all layers of the network, which normalizes across the spatial dimension to generate qualitatively superior images during training and at test time.

  • G1的输入是LR的灰度图的邻近插值和LR的灰度图的canny 边缘图的邻近插值，输出是对HR的灰度图的canny 边缘图的预测
  • 
  • G1比常规G加一个${\mathcal{L}}_{FM}$无法用perceptual loss因为VGG不是训练在边缘图上的，没有作用
  • 还使用了GAN的SN技术
  • G2的输入是LR的fixed fraction conv和G1的输出，就是针尖的upsample，卷积核长这样：
  •