Channel attention and spatial attention module

Channel attention module

The purpose of using the channel attention module ： In order to make the input image more meaningful , The general understanding is , Calculate the input image through the network The importance of each channel （ The weight ）, That is, pay more attention to which channels contain key information , Pay less attention to channels with little important information , So as to achieve Improve feature representation Purpose .

In short ： Attention mechanism can correct the characteristics , The corrected features can retain valuable features , Eliminate worthless features .

Channel attention mechanism steps ：

• extrusion （Squeeze） The input image

  Compress the spatial dimension of the input feature graph , This step can be done by Global average pooling （GAP） and Global maximum pooling （GMP）（ The effect of global average pooling is relatively better than that of maximum pooling ）, Through this step .HxWxC The input image is compressed into 1x1x``C Channel descriptor for . Enter the following formula as SxSxB Of feature map：
  
  ​ Compress the global spatial information into the channel descriptor , It not only reduces the network parameters , It can also prevent over fitting .

• excitation Channel descriptor

  This step is mainly to send the channel descriptor obtained in the previous step to two fully connected networks , Get the attention weight matrix , Then multiply with the original image to obtain the calibrated attention feature map .

Spatial attention module

The purpose of using channel attention ： Find the key information in map Where is the most , yes Supplement to channel attention , Simply speaking , Channel attention is to find out which channel has important information , Spatial attention is based on this , Based on the direction of the channel , find Which location The most information gathered .

Spatial attention steps ：

• Apply average pooling and maximum pooling operations along the channel axis , Then connect them to generate a valid feature descriptor .

  Be careful ： The pooling operation is carried out along the channel axis , That is, the values between different channels are compared during each pooling , Instead of values in different areas of the same channel .

• The feature descriptor is sent into a convolution network for convolution , The final spatial attention feature map is obtained through the activation function .

say concretely , Use two pooling Operations are aggregated into one feature map Channel information , Generate two 2D chart : Fsavg The size is 1×H×W,Fsmax The size is 1×H×W.σ Express sigmoid function ,f7×7 Indicates that the size of a filter is 7×7 Convolution of .

Reference link ：

（1）https://blog.csdn.net/u011984148/article/details/109475440

（2）https://zhuanlan.zhihu.com/p/334349672

（3）https://zhuanlan.zhihu.com/p/101590167

349672

（3）https://zhuanlan.zhihu.com/p/101590167