[Images] 🧹 Deprecate imROF in favor of solve_ROF_PD

Because the boundary handling of divergence has changed, this is not an
identical implementation of `imROF`. Thus old test codes are removed in this
commit. Instead, ImageFiltering has more comprehensive test codes.

See also: JuliaImages/ImageFiltering.jl#233

This commit bumps ImageFiltering compatibility to at least v0.7.1

Author: johnnychen94

packages/Images/Project.toml

@@ -37,7 +37,7 @@ ImageBase = "0.1.5"
 ImageContrastAdjustment = "0.3.3"
 ImageCore = "0.9.3"
 ImageDistances = "0.2.5"
-ImageFiltering = "0.6.3"
+ImageFiltering = "0.7.1"
 ImageIO = "0.0.1, 0.3, 0.4, 0.5"
 ImageMagick = "0.7.5, 1"
 ImageMetadata = "0.9.7"

packages/Images/src/Images.jl

@@ -155,7 +155,6 @@ export
     findlocalminima,
     imgaussiannoise,
     imlineardiffusion,
-    imROF,
     otsu_threshold,
     yen_threshold,
 

packages/Images/src/algorithms.jl

@@ -301,49 +301,6 @@ end
 imgaussiannoise(img::AbstractArray{T}, variance::Number) where {T} = imgaussiannoise(img, variance, 0)
 imgaussiannoise(img::AbstractArray{T}) where {T} = imgaussiannoise(img, 0.01, 0)
 
-"""
-```
-imgr = imROF(img, λ, iterations)
-```
-
-Perform Rudin-Osher-Fatemi (ROF) filtering, more commonly known as Total
-Variation (TV) denoising or TV regularization. `λ` is the regularization
-coefficient for the derivative, and `iterations` is the number of relaxation
-iterations taken. 2d only.
-
-See https://en.wikipedia.org/wiki/Total_variation_denoising and
-Chambolle, A. (2004). "An algorithm for total variation minimization and applications".
-    Journal of Mathematical Imaging and Vision. 20: 89–97
-"""
-function imROF(img::AbstractMatrix{T}, λ::Number, iterations::Integer) where T<:NumberLike
-    # Total Variation regularized image denoising using the primal dual algorithm
-    # Also called Rudin Osher Fatemi (ROF) model
-    # λ: regularization parameter
-    s1, s2 = size(img)
-    p = zeros(T, s1, s2, 2)
-    # This iterates Eq. (9) of the Chambolle citation
-    u = similar(img)
-    τ = 1/4   # see 2nd remark after proof of Theorem 3.1.
-    for i = 1:iterations
-        div_p = ImageBase.FiniteDiff.fdiv(view(p, :, :, 1), view(p, :, :, 2))
-        u .= img - λ*div_p # multiply term inside ∇ by -λ. Thm. 3.1 relates this to u via Eq. 7.
-        grad_u = cat(ImageBase.fdiff(u, dims=1, boundary=:zero), ImageBase.fdiff(u, dims=2, boundary=:zero), dims=3)
-        grad_u_mag = sqrt.(sum(abs2, grad_u, dims=3))
-        p .= (p .- (τ/λ).*grad_u)./(1 .+ (τ/λ).*grad_u_mag)
-    end
-    return u
-end
-
-# ROF Model for color images
-function imROF(img::AbstractMatrix{<:Color}, λ::Number, iterations::Integer)
-    out = similar(img)
-    imgc = channelview(img)
-    outc = channelview(out)
-    for chan = 1:size(imgc, 1)
-        outc[chan, :, :] = imROF(imgc[chan, :, :], λ, iterations)
-    end
-    out
-end
 
 # morphological operations for ImageMeta
 dilate(img::ImageMeta, region=coords_spatial(img)) = shareproperties(img, dilate!(copy(arraydata(img)), region))

packages/Images/src/deprecations.jl

@@ -1081,6 +1081,7 @@ end
 @deprecate backdiffx(X) ImageBase.FiniteDiff.fdiff(X, dims=2, rev=true, boundary=:zero)
 @deprecate backdiffy(X) ImageBase.FiniteDiff.fdiff(X, dims=1, rev=true, boundary=:zero)
 @deprecate div(A::AbstractArray{<:Any,3}) ImageBase.FiniteDiff.fdiv(view(A, :, :, 1), view(A, :, :, 2)) false
+@deprecate imROF(img::AbstractMatrix, λ::Number, iterations::Integer) ImageFiltering.Models.solve_ROF_PD(img, λ, iterations)
 
 
 # This is now replaced by ImageTransformations and Interpolations

packages/Images/test/algorithms.jl

@@ -467,24 +467,6 @@ using ImageBase.FiniteDiff: fdiff
         img = zeros(Gray{Float64},10,10,3)
         @test yen_threshold(img) == 0
     end
-
-    @testset "imROF" begin
-        img = [0.1 0.2 0.1 0.8 0.9 0.7;
-               0.2 0.1 0.1 0.8 0.1 0.8;
-               0.1 0.2 0.1 0.7 0.9 0.8]
-        # # Ground truth
-        # using Optim
-        # diff1(u) = [u[2:end,:]; u[end:end,:]] - u
-        # diff2(u) = [u[:,2:end] u[:,end:end]] - u
-        # obj(Avec, img, λ) = (A = reshape(Avec, size(img)); sum(abs2, A - img)/2 + λ*sum(sqrt.(diff1(A).^2 + diff2(A).^2)))
-        # res = optimize(v->obj(v, img, 0.2), vec(img); iterations=10^4)
-        # imgtv = reshape(res.minimizer, size(img))
-        target = [fill(0.2, (3,3)) fill(0.656, (3,3))]
-        @test all(map((x,y)->isapprox(x, y, atol=0.001), imROF(img, 0.2, 1000), target))
-        imgc = colorview(RGB, img, img, img)
-        targetc = colorview(RGB, target, target, target)
-        @test all(map((x,y)->isapprox(x, y, atol=0.001), channelview(imROF(imgc, 0.2, 1000)), channelview(targetc)))
-    end
 end
 
 nothing