Enhanced Deep Residual Networks for Single Image Super-Resolution（EDSR） 论文笔记

Enhanced Deep Residual Networks for Single Image Super-Resolution

• Abstract:

1. Use Residual Networks but optimize it by analyzing and removing unnecessary modules
2. Do some modification on training multi-scale training
3. PSNR and SSIM

• Related work:

1. Interpolation base on sample theory:
   low resolution image is down sample of the high resolution image
2. Adopt nature image statistic to problem (?)
3. Neighbor embedding (machine learning method)
4. Sparse coding （machine learning method）
5. Deep Neural Networks:

• SRCNN 香港中文大学，汤晓鸥:
  3 convolutional layers:
  feature extraction --> non-linear mapping --> reconstruct

• Residual Network for SR:
  To get deeper network and superior performance

• Use Recursive Convolutional Network:
  《Deeply-Recursive Convolutional Network for Image Super-Resolution》
  alleviate the burden of carrying identity information in the super-resolution network???.

• Encode and Decode structure with skip connections:
  they said it provide fast and improved convergence

6. Training strategy:

1. Many method do the upsample operation via bicubic Interpolation before the image feed to the Network
2. In VDSR, they train a single Network using multi-scale images 2X 4X 8X so the network can deal with multi-scale problem.
3. Geometric Self-ensemble, do some transform operation to the lr images as a set and then feed them to the network（个人理解是生成多个针对于不同角度变换lr的学习模型，然后将这些模型给一些权重，集成成一个更具有鲁棒性的学习器）
   

7. Loss function: MSE or L2 loss, SSIM

• Propose method:
  remove BN: they get rid of range flexibility from
  networks by normalizing the features ???
• 总体来说这篇文章是在resnet网络的基础上对网络的多余部分进行了精简，例如删掉BN层等操作，并且使用了以前提到的multi-scale训练策略来进行实验。