Second-order Attention Network for Single Image Super-Resolution

• Contributions:
  • Adding second order Attention to learn feature interdependencies
  • (It is very similar to RCAN that is used to first order Attention)

  

  Fig.1: SANSR blocks

Core parts of the system design:

1. Shallow feature extraction
   A basic convolution network is used to extract shallow features.

2. Deep feature extraction
   
   • Region-Level Non-Local Network(RL-NL)
     Non-Local Neural Network is used
   • Share Source Residual Group(SSRG)
     • Local Source Residual Attention Group(LSRAG)
       • Residual blocks
       • Second Order Channel Attention(SOCA)
     • Share Source Skip Connection(SSC)
3. Upscale
   ESPCNN upscaling method is used. It uses 2 convolution layer to extract features and image is scaled with phase shift method. (This should be reviewed too)
4. Reconstruction
   Reconstruction of RGB image from upscale network.

Fig.2: SSRG-part1

Fig.3: SSRG-part2

Fig.4: SSRG-part3