CV convolution neural network

Convolution kernel

filter , Templates
Weighted summation

another F For image ,H Is the convolution kernel ,F And H The convolution is written as R=F*H
$$R_{ij}=\sum _{u,v}{H}_{i-u,j-v}{F}_{u,v}$$

Border filling

If boundary filling is not considered, the output will become smaller

• 0 fill （ Commonly used ）
• The tensile
• Mirror image

The bottleneck of fully connected neural networks

Too many parameters
Not suitable for processing images , It is only suitable for processing feature vectors and small images

Convolution kernel in convolution network

A convolution kernel corresponds to a characteristic response
Multiple convolution kernels correspond to multiple characteristic responses

The number of channels of convolution kernel is equal to the number of upper convolution kernel

Different positions of the characteristic response graph represent the responses of different positions of input to the same convolution kernel

Convolution has a step size

The size after convolution
$$\begin{gathered} W_2=(W_1-F)/S+1 \\ {H}_2=(H_1-F)/S+1 \end{gathered}$$
You can fill the boundary

pooling layer

The feature corresponding graph shall be carried out independently , Reduce the width and height of the corresponding graph , Reduce the amount of parameters , Save computing resources , Prevent over fitting , Expand the feeling field

Common pooling

• Maximum pooling —— Non maximum suppression
• The average pooling

Image enhancement

Existing problems —— Too few samples lead to over fitting
Image enhancement —— Generate more training data from existing samples
The goal is —— Observe more data , Enhance generalization performance
Method —— Mirror image , The zoom , Cutout , Color jitter , The tensile , distortion …