How to transform an image into a probability density function?

Question

I'd like to use the Metropolis algorithm to randomly generate points with density according to the brightness of an image. I just need to transform a binary image into a pdf function to use this answer.

First attempt:

img = ColorNegate@ColorConvert[ImageResize[lena[], 50], "Grayscale"];
dims = ImageDimensions[img];
data = Flatten[
   Table[{{i, j}, PixelValue[img, {i, j}]}, {i, dims[[1]]}, {j, 
     dims[[2]]}], 1];
f = Interpolation[data]

ContourPlot[f[x, y], {x, 1, dims[[1]]}, {y, 1, dims[[2]]}]

Metropolis /: 
  Random`DistributionVector[
   Metropolis[pdf_, u0_, s_: 1, n0_: 100, chains_: 200], n_Integer, 
   prec_?Positive] := 
  Module[{u, du, p, p1, accept, cpdf}, 
   cpdf = Compile @@ {{#, _Real} & /@ #, pdf @@ #, 
        RuntimeAttributes -> {Listable}, RuntimeOptions -> "Speed"} &[
     Unique["x", Temporary] & /@ u0];
   u = ConstantArray[u0, chains];
   p = cpdf @@ Transpose[u];
   (Join @@ 
      Table[du = 
        RandomVariate[
         NormalDistribution[0, s], {chains, Length[u0]}];
       p1 = cpdf @@ Transpose[u + du];
       accept = UnitStep[p1/p - RandomReal[{0, 1}, chains]];
       p += (p1 - p) accept;
       u += du accept, {Ceiling[(n0 + n)/chains]}])[[n0 + 1 ;; 
      n0 + n]]];


p = RandomVariate[Metropolis[f, {25, 25}], 30000];
ListPlot[p, AspectRatio -> Automatic]

Why not use HistogramDistribution[] or SmoothKernelDistribution[]? Related to this, look up the docs for ImageHistogram[]. — J. M.'s missing motivation, Aug 14 '15 at 15:35
@Guesswhoitis. There are no examples in the Docs for a 3d histogram from a black and white image... — M.R., Aug 14 '15 at 15:58
If the intensity represents the PDF magnitude and not the raw data then HistogramDistribution and similar are not what you want. — rhermans, Aug 14 '15 at 16:03

Niki Estner · Accepted Answer · 2015-08-14T18:43:50.063

If you don't care about the algorithm and only want to sample points with density according to image brightness, you could just use RandomChoice:

using a test image that looks a little bit like a PDF:

img = Image[
   Rescale[Array[
     Sin[#1^2]*Cos[#2 + Sin[#1/5]] + Exp[-(#1^2 + #2^2)/2] &, {512, 
      512}, {{-2., 4.}, {-3., 3.}}]]];

I can then sample random pixel indices weighted by the corresponding pixel brightness:

weights = Flatten[ImageData[img]];    
sample = RandomChoice[weights -> Range[Length[weights]], 10000];

And convert indices back to coordinates:

{w, h} = ImageDimensions[img];    
pts = Transpose[{Mod[sample, w], h - Floor[sample/N[w]]}];

(this is blazingly fast. Sampling 10^6 points takes about 0.2 seconds.)

Show[img, Graphics[{Red, PointSize[Small], Opacity[.5], Point[pts]}]]

ADD: If you really want to use the MH algorithm, why not use ImageValue directly, instead of creating an interpolation function?

pdf = Function[{x, y}, ImageValue[img, {x, y}]];
ContourPlot[pdf[x, y], {x, 0, 511}, {y, 0, 511}]

How to transform an image into a probability density function?

1 Answers1