CNN--Introduction to each layer

卷积层Convolutional

给定卷积核,Where the convolution kernel is applied to the corresponding dimension of the input image features,计算乘积,The figure below shows the step sizestride=1为例,Get the final output layer features.

假设输入大小为(H,W),滤波器大小为(FH,FW),输出大小为(OH,OW),填充padding为P,步幅stride为S
$$\begin{gathered} OH=\frac{H+2P-FH}{S}+1 \\ OW=\frac{W+2P-FW}{S}+1 \end{gathered}$$
For multi-channel convolution calculations,Similar to the single channel calculation method,Just add the calculated values ​​for each channel at the end

Convolution operation for multiple convolution kernels：

填充Padding

after each convolution,The output dimension is reduced,according to the size of the input image,The dimensions of the output image may become too small after several rounds of convolution,At the same time, the pixels on the edge are less than the pixels in the middle,This also ignores part of the image data,为了解决这个问题,PaddingBy padding data at the edges,To achieve the effect of keeping the input and output image dimensions consistent.

池化层Pooling

Pooling layer to reduce the size of the special row data,And make some feature detection more robust.如果使用一个$4\times 4$的矩阵,Max池化层和MeanThe result after the pooling layer is processed separately is shown below,这个过程很简单.在示例中,filter是2×2,stride为 2,So divide the input into four parts $2\times 2$ 的子区域,Max和MeanThen, the maximum and mean values ​​of the corresponding sub-regions are output respectively

的子区域,Max和MeanThen, the maximum and mean values ​​of the corresponding sub-regions are output respectively

The above picture is from《深度学习入门：基于Python的理论与实现》and web images