对卷积神经网络 3D 或多光谱图像中 2D、3D 卷积的理解混淆

Question

有人请解释使用 3D 或多光谱图像的 CNN（深度学习）中 2D 和 3D 卷积之间的区别吗？

Answer 1

all convolution operation has padding for maintaining output size [w, h]

• 2D Convolution

  input size = [w, h]

  conv filter = [n, n]

  output size = [w, h]

• 3D convolution

  input size = [w, h, c]

  conv filter = [n, n, d]

  output size = [w, h, c']

• 2D convolution for 3D input (usually used this form)

  input size = [w, h, c]

  conv filter = [n, n, c]

  output size = [w, h]

  if you want to get output size with [w, h, c'], you need c' times operation.