The relationship between the number of channels and the number of convolution kernels of convolution neural networks.

problem description

if you convolution a three-channel picture, the number of convolution kernels is 10. Since the final output of 10 feature map, does a convolution kernel have three channels? each convolution checks the image convolution of the three channels and then adds up to get a feature map?.

do I understand right?

Aug.16,2021

your understanding is correct!
see the following figure:

  1. there are a total of two convolution cores in the above figure, W0 and W1, both of which are 3x3x3. The front 3x3 is the convolution core size, and the last 3 is the channel of the input feature map. If the form in the convolution function of tensorflow is 3x3x3x2, the last 2 is the number of channels of the output feature map, which is also the number of convolution cores.
  2. take w0 as an example, the numerical values of each position of the three channels in the blue box A = [0pje 0je 0je 0je 1], [0pr 0je 0ret 1m 1m 1m 0J 0J 0 2], [0Me 0Me 0Me 02pm 0J 0J 0]} multiplied by the values of each position of each channel of the pink box (convolutional core) W0 = [[1mlmlmlle 1lim 1p 1p 1p 1pm]. Finally, add the offset bread1, and you will get the value in the green box. AW0+b=output0
  3. Welcome to visit: hometown.group many personal knowledge summary, if the previous URL can not be accessed, may be changed to learncv.cn, ready to change the domain name.
Menu