0%

(CVPR 2018) xUnit:Learning a Spatial Activation Function for Efficient Image Restoration

Posted on In , ,

Keyword [xUnit] [Activation Unit] [GCN]

Kligvasser I, Rott Shaham T, Michaeli T. xUnit: learning a spatial activation function for efficient image restoration[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 2433-2442.


1. Overview

目前

  • 0% parameters in Activation unit (nonlinearities), Conv (spatial processing)
  • Network deeper and deeper

因此，论文提出xUnit结构

  • learnable nonlinear function with spatial connection
  • reduce layer
  • used in low-level task (denoising, de-raining, super resolution)


  • ELUs
  • SRCNN
  • VDSR
  • SRResNet
  • EDSR
  • ESPCN
  • Binarized Neural Networks
  • Deep Detail Network
  • DehazeNet
  • MobileNet

1.2. 模型



Conv+Activation形式如下：



1.2.1. ReLU



  • o表示点乘，定义0/0=0.

1.2.2. xUnit



H_k为depth-wise convolution.
对于d-channel输入，d-channel输出而言，计算复杂度为

  • 标准Conv. rxrxdxd
  • depth-wise Conv. rxrxd


2. Experiments

2.1. Compare




2.2. 特征图可视化

ReLU丢弃了大部分特征图（黑色），而xUnit大部分特征图都参与后续计算（白色）。



2.3. Denoising




2.4. De-raining



(PSNR) 28.94 VS 31.17.

2.5. Super Resolution