Enhanced Deep Residual Networks for Single Image Super-Resolution

[harshraj22]

Enhanced Deep Residual Networks for Single Image Super-Resolution 📓

CVPRW 2017

Architectural details:

• Built on the top of generator of pre-existing SRResNet model. The authors reduce the number of parameters by removing the unnecessary modules, without affecting the accuracy much.
• The architectural blocks are almost similar for cases when the image is upsampled by x2, x3, x4. This allows them to share weights for most blocks from a pretrained network across the sampling.
• They use L1 loss as the loss function for training the networks.
• The benchmark their results on various dataset, and achieve good results on Div2k dataset.
• They consider the following 2 metric for comparison of their result:
  - PSNR
  - SSIM

Dataset Details:

• The model was trained on DIV2K dataset containing 800 images on train split.
• For training, they use the RGB input patches of size 48×48 from LR image with the corresponding HR patches.
• They augment the training data with random horizontal flips and 90 rotations and pre-process all the images by subtracting the mean RGB value of the DIV2K dataset