Calculation of feature map size, receptive field, and parameter amount after convolution operation

1. The size of the output feature map

• If dilated convolution is included, that is, when the dilation rate is not 1:

2. Calculation of receptive fields

• For example, after the image undergoes two 3*3 convolutions with a stride of 2, the receptive field is:

conv1: F = (1 - 1) * 2 + 3 = 3
conv2: F = (3 - 1) * 2 + 3 = 7

3. Parameters of the convolution kernel

4. Computational amount of standard convolution and depth-separable convolution

Input feature map: $C_{in}\times H\times W$
standard convolution:$C_{out}$个 $K\times K\times C_{in}$The convolution kernel
is depth separable convolution: 1 $K\times K\times C_{in}$Depth convolution + $C_{out}$个$1\times 1\times C_{in}$Point-wise convolution
output feature map: $C_{out}\times H^{\prime }\times W^{\prime }$

• The calculation amount of standard convolution is $K\times K\times H^{\prime }\times W^{\prime }\times C_{in}\times C_{out}$
• The calculation amount of depth-separable convolution is $K\times K\times H^{\prime }\times W^{\prime }\times C_{in}+1\times 1\times H^{\prime }\times W^{\prime }\times C_{in}\times C_{out}$

Reference links:
1) https://blog.csdn.net/DBlueink/article/details/114878888
2) https://blog.csdn.net/qq_36412570/article/details/115347200