Cleaner way for applying conv2 to RGB image

question

#1

I am looking to apply conv2 on a RGB image. However the two solutions I arrived at don’t look very good. Questions:

  1. Is there any a better way to apply conv2 along a given dimension of img and return the result?
  2. Is there a cleaner way of writing the second solution below? I need to concatenate the three nested multidimensional arrays in channels along a new dimension in a new 3xHxW Array.

I know I can use the api from Images for filtering, but this is just an exercise in Julia syntax.

img = rand(3, 10, 16)
g = fill!(rand(3, 3), 0.001)
g[2, 2] = 1

# convolve each channel
channels = [conv2(img[i, :, :], g) for i in 1:3]

# solution no 1
new_size = (3, size(channels[1])...)
result = zeros(new_size)
for i in 1:3
  result[i, :, :] = channels[i]
end

# solution no 2
new_size = (size(channels[1])..., 3)
result = permutedims(reshape(hcat(channels...), new_size), [3, 1, 2])

Thank you!