An Attention Module for Convolutional Neural Networks论文解读

An Attention Module for Convolutional Neural Networks

Our contributions are summarized as follows：
We point out and analyze two ignored problems of the current attentional activations-based models: the approximation problem and the insufficient capacity problem of the attention maps. To address the two problems together, we originally propose an attention module by developing an AW convolution,where the shape of the attention maps matches that of the weights instead of the activations.

主要解决现有基于激活的注意力模型中的两个问题：近似问题和能力不足问题而提出AW-Convolution模块.

• The reason of Approximation problem of the attention maps:

Instead of directly computing the three-dimensional attention map(N is omitted, otherwise the attention maps are of four dimensions.) ，all the current attentional activations-based models produce the attention maps separately into the channel attention maps and spatial attention maps, which leads to the approximation problem of attention maps.

目前所有的基于激活的注意力模型都是使用通道注意力映射和空间注意力映射，因此会产生近似问题。

• The reason of Insufficient capacity problem of the attention maps:

To compute different channels of the output activations of the convolution, the input activations are constrained to be recalibrated by the same attention maps, which indicates the insufficient capacity of the attention maps

输入激活的能力会因为相同的注意力映射的重复计算而被限制，因此会产生注意力映射能力不足的问题。

• 因此，提出了AW-Convolution模块解决这两者的问题：

模块如图所示：

其中:

.
Avgpool2d聚合来自整个空间范围的响应，并将它们嵌入到A0中，Pointconv1和Pointconv2之后Relu重新分配池中的信息，以捕获通道和空间空间之间的动态和非线性依赖关系。

F2如此改动可以引入残差结构

AW-Convolution(I,AK),就是A和卷积核K进行点乘，后再使用点乘后的卷积核AK对I进行卷积。（个人理解，如有错误，望指出）

将AW-Convolution模块加入模型的例子：

心得：
该模块最主要就是将输入进行注意力映射，再将输入的注意力映射和权重进行匹配获取想要的注意力位置。