Normall convolution

approximate complexity = out_dim x kernel x conv = 4 x 9 x conv. (conv = in_dim x in_pixels)

image

Separable convolution

Separable conv = Depth-wise conv + Point-wise conv

approximate complexity = kenel x conv + out_dim x conv = (9 + 4) x conv. (conv = in_dim x in_pixels(

image image

Animation and videos

(Youtube video) Groups, Depthwise, and Depthwise-Separable Convolution (Neural Networks)

Animation page