DEGREE.md

Short introduction

Main contributions

• integrate edge priors
• new framework to integrate prior knowledge
• recurrent residual structrue

Architecture

Dense compression unints

Recurrence

Loss

• HR image and HR edges.
• Discriminator loss: MSE

Training strategy

• Merging preditions from different sub-band
• Predict sub-band from acculmulating bands
• Randomly select one of the scales s to avoid mixing batch statistics

Experiments

• Dataset: 91, only on u channel, other channel using bicubic upsampling
• Evaluation metric: PSNR and perceptual
• Patchsie: 33 × 33 with rotation and flip as agumentation
• SGD, learning rate from 0.0001, 270 epochs, 0.9 momentum

Final summary

Pros:

• Making use of edge info to improve high frequency detials
• recurrence and residual structure for better performance

Cons:

• No detial about the hand-crafted edge gt obtaining

Tips:

• SR is about the enhacement of info from different frequencies