0%

(CVPR 2016) Deeply-recursive convolutional network for image super-resolution

Kim J, Kwon Lee J, Mu Lee K. Deeply-recursive convolutional network for image super-resolution[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 1637-1645.



1. Overview


1.1. Motivation

  • stacking Conv to increase receptive filed lead to more parameter
  • pooling discard pixel-infomation

In this paper, it proposed DRCN (deeply-recursive convolutional network)

  • 16 recursion
  • recursive-supervision and skip-connections to solve gradient problem



2. Methods


2.1. Baseline




2.2. Advance Model



  • D predictions (shared net to predict) are simultaneously supervised
  • use all D predictions to compute the final output
  • skip connections.
    • LR and HR are similar
  • weighted average of all intermediate predictions (learned parameter)


2.3. Loss Function

  • intermediate output



  • final output



  • final loss function





3. Experiments


3.1. Comparison



3.2. Ablation Study