Packs the image into a [1, 3, height, width] float tensor: (value - mean) * scale per channel value.
[1, 3, height, width]
(value - mean) * scale
Common recipes: raw 0..255 (mean = 0, scale = 1), or [-1, 1] (mean = 127.5, scale = 1/127.5).
mean = 0, scale = 1
[-1, 1]
mean = 127.5, scale = 1/127.5