Second-order Attention Network for Single Image Super-Resolution

• Second-order Attention Network for Single Image Super-Resolution

• 性能：
• 
• SSRG是提出来的一个新的backbone，看图就行，每个块就是$F_g=W_{SSC}(F_0)+H_g(F_g-1)$其中$W_{SSC}$就是共享的权重
• RL-NL把feature divide为k*k个grid，对每个grid做non-local operations，至于什么是non-local operation，看这个网站https://zhuanlan.zhihu.com/p/33345791
• LSRAG就是提SSRG每个小块的细节，看图就行
• SOCA倒是个复杂的东西：
  • 先把前一层进来的feature map$F=[f_1,\cdot \cdot \cdot ,f_C]$ reshape 成(C，s)的shape，其中s = H*W，设reshape后的矩阵为X
  • 然后将X normalize，具体而言就是$\Sigma =X\overline{I}{X}^T$,这时Σ 又可以分解为$\Sigma =U\Lambda U^T$ 其中Λ为对角阵，$\hat{Y}={\Sigma }^{\alpha }=U{\Lambda }^{\alpha }{U}^T$当α小于1时可以将Λ中的元素拉向1，文中设为0.5，注意这时$\hat{Y}$的shape为(C,C)
  • $\hat{Y}=[y_1,\cdot \cdot \cdot ,y_C],z_c=H_{GCP}(y_c)=\frac{1}{C}\sum _i^C{y}_c(i)，w=f(W_U\delta (W_{D}z))$其中$z_c$为每一层channel的cov（(1,C)）经过global channel average pooling之后的输出(scalar)，而z将z_c统一为一个(C,1)的向量, δ为relu，f为sigmoid，这里得到的w就是attention的channel wise的权重：$\widehat{f_c}=w_c\cdot f_c$
• 其它一些细节参数如下

We set LSRAG number as G = 20 in the SSRG structure, and embed RL-NL modules (k = 2) at the head and tail of SSRG. In each LSRAG, we use m = 10 residual blocks plus single SOCA module at the tail. In SOCA module, we use 1 × 1 convolution filter with reduction ratio r = 16. For other convolution filter outside SOCA, the size and number of filter are set as 3 × 3 and C = 64, respectively. For upscale part H↑(·), we follow the works in [20, 39] and apply ESPCNN [28] to upscale the deep features, followed by one final convolution layer with three filters to produce color images (RGB channels).