0%

(CVPR 2018) Residual-Guide Feature Fusion Network for Single Image Deraining

Fan Z, Wu H, Fu X, et al. Residual-Guide Feature Fusion Network for Single Image Deraining[J]. arXiv preprint arXiv:1804.07493, 2018.



1. Overview


1.1. Motivation

  • wasteful to utilize a resource-hungry model to meet all kinds of demands for rain streak removal task
  • cascaded architecture may lost valuable intermediate reconstruction features


In this paper, it proposed ResGuidedNet

  • residual generated from shallower blocks to guide deeper blocks
  • coarse to fine estimation
  • recursive convolutional
  • apply to low-level and high-level



2. Methods




  • recursive block. Conv+LeakyReLU, 5 recursive time
  • ensemble all intermediate results to predict final output



2.1. Loss Function

  • L2 + SSIM
  • for block_k


  1. g. SSIM function
  2. λ=1
  • M + 1 loss function




3. Experiments


3.1. Comparison



3.2. Ablation Study



3.3. Time