Behavior of limits

[lucasb-eyer]

Continuing the discussion at #100 because it got off-topic.

limits still leaves a bitter taste, because of behavior like the following:

julia> img = convert(Image, Uint8[1 240; 10 128])
Gray Image with:
  data: 2x2 Array{Uint8,2}
  properties:
    timedim: 0
    colorspace: Gray
    colordim: 0
    spatialorder:  y x
    pixelspacing:  1.0 1.0
    limits: (0x00,0xff)

julia> img * 2
Gray Image with:
  data: 2x2 Array{Uint8,2}
  properties:
    timedim: 0
    colorspace: Gray
    colordim: 0
    spatialorder:  y x
    pixelspacing:  1.0 1.0
    limits: (0,510)

julia> data(img * 2)
2x2 Array{Uint8,2}:
 0x02  0xe0
 0x14  0x00

I'm not sure on how to best handle this, esp. in the case of Uint16 images filling only 12 bits for example.

[timholy]

Yikes, I didn't know about that. In my personal view, the problem is not limits per se, the problem is the inconsistency in the element type between how limits is treated and Julia's rules for array arithmetic.

See if you like the solution in #102. What else?

[lucasb-eyer]

Sweet implementation :)

I cannot really put my finger on what exactly it is that leaves a bitter taste. I keep wondering whether it's sensible/worth the effort to follow OpenCV and use saturation-arithmetics instead of overflow-arithmetics. And if so, whether it should clamp at the typemin/typemax or at the limit. I'm thinking of a 12bit image stored in an Image{Uint16} for example, should img*2 clamp to 4096 or raise the limit to 8092, but arbitrarily stop raising the limit when it reaches 65536?

[timholy]

I haven't implemented this yet, but here's a mock-up that shows another way of thinking about this (carefully note the element types, and if you've used SIUnits the basic concepts should be familiar):

julia> img = imread("img.nrrd")
Gray Image with:
  data: 512x512 Array{Count{Uint16},2}
  properties:
    colorspace: Gray
    spatialorder:  x y
    limits: (0x0000 Count,0x0fff Count)
    gain: 3.8 Photons/Count

julia> imgp = img["gain"] * img
Gray Image with:
  data: 512x512 Array{Photons{Float64},2}
  properties:
    colorspace: Gray
    spatialorder:  x y
    limits: (0.0 Photons,15561.0 Photons)
    gain: 3.8 Photons/Count

Once you attach physical units to the intensities, it seems clearer to me that the limits should scale along with the values.

[lucasb-eyer]

The taste is gone, thanks.