Personal understanding of convolution calculation process of convolution neural network

1. The position of convolution calculation in neural network

Let's take a simple example to illustrate .
The above figure is a brief process diagram of convolutional neural network drawn by myself . The position of convolution is from Input To Feature The process of .

Why is it called convolutional neural network ？
Because every convolution calculation is similar to a neuron , Collect previous results , Then output a result through an activation function .

h_wb(x)=X₀W₀ + X₁W₁ + X₂W₂ + 1

The calculation process is from the first figure Input To Feature A calculation of .

2. The main function of convolution calculation

The main function of convolution is to extract features . The training of network is mainly to find the best parameter combination of convolution , Make the recognition accuracy the highest . In the training process, the value of the convolution kernel is repeatedly modified by comparing the difference between the category calculated by the current convolution kernel and the real category , Slowly achieve the best accuracy .

3. The process of convolution calculation

The whole picture may be 800x600 Pixel size , But the convolution kernel may only be 3x3 size , This requires sliding to traverse the entire image pixel value . This is the convolution stride. In the process of calculation, it will be found that some edges cannot be obtained during calculation , This involves padding, You can refer to another blog https://blog.csdn.net/itlilyer/article/details/107288176.

The following figure is borrowed from other articles , It clearly describes the calculation process of convolution .