Should we use N0f8 to store image?

We use N0f8 to store image data in default.

But I wonder if this does correctly.

using Images

# my custom function
function down_intensity!(img::Array{Gray{T}}, bit) where T
    gray_level = range(0, stop = 1, length= Int(2^ bit +1))
    inds = [gray_level[i] .< img .<= gray_level[i+1] 
        for i in 1:length(gray_level)-1]
    # This for-loop is slow
    for i in 1:length(gray_level)-1
        img[inds[i]] = fill(gray_level[i],length(findall(inds[i])))
    img = Gray{T}.(img/maximum(img))
    return img
function down_intensity(img::Array{Gray{T}}, bit) where T
    dest_img = copy(img)
    down_intensity!(dest_img, bit)

graybar = Gray{Float32}.(transpose(repeat(range(0,stop=1,length=512),1,50)))

bits = 4:-1:1
imgs_builtin = [Gray{N0f8}.(Gray{Normed{UInt8, bit}}.(graybar)) for bit in bits]
imgs_custom = [down_intensity(graybar, bit) for bit in bits]

img_diffs = colorview(Gray,hcat(vcat(imgs_builtin...),vcat(imgs_custom...)))

The left column shows results of the builtin way, the right column shows that of my way. As you can see, the left column is not spanned uniformly; there’s more on the center, less on both sides.


I find some note saying pixel at i, j is a sensor that integrates the intensity over an area spanning i±0.5, j±0.5 (this is a good model of how a camera pixel actually works). In this opinion, N0f8 seems reasonable.