From what I know, separable filters will reduce the cost of computation.
If the image has M x M pixels and the filter kernel has size N x N, it said that the cost of convolution with a non-separable filter is $M^2 \times N^2$. If we can decompose the N x N filter into one vertical N (column) and one horizontal N (row), what will be the cost of the convolution with a separable filter?
I am expecting the answer to be 2 * M^2 * N (?) However I will need more clarification for that. Any explanation will be greatly appreciated.