Normally you surround all the holes with planar curve that doesn't touch any of the holes, then pass it and all the holes (flattened) to a 'Boundary' surface.  However, it usually doesn't work (for me) with that many holes; the computer tries forever and I eventually have to kill GH/Rhino.

This is a method that works well. It's a complete re-write with various features to tune the results, including scale, resolution and the number of perforated sub-surfaces (more reduces the number of holes on each so speed increases). Zero radius circles are skipped.

The image dimensions in pixels are defined in small panels (X=485, Y=759) and used to calculate height/width ratio. That is used to define height based on the 'X' slider (500), which defines width overall.

The 'cell size' slider is also in units and determines resolution indirectly. For any given X value, increasing 'cell size' reduces the number of grid cells (resolution) and vice-versa.

Independent of other parameters, 'Isotrim (SubSrf)' splits the base surface into sub-surfaces, onto which the circles are projected. The 'SrfSplit' does the heavy lifting (can be SLOW!) and finally, 'Sort' is used to select the resulting surfaces that contain the holes.

Benchmarks:

X = 500, cell size = 10, 3161 circles
nine subsurfaces: 'SrfSplit' = 6.6 minutes, 'Project' = 13 secs.
16 subsurfaces: 'SrfSplit' = 2.3 minutes, 'Project' = 17 secs.

X = 500, cell size = 5, 12542 circles  (shown)
35 surfaces: 'SrfSplit' = 30.6 minutes, 'Project' = 57 secs.

As noted before, a very long-standing, well-known bug in Grasshopper fails to save the Image Sampler component when I save a copy of your file. Very annoying, but there is a work-around. Copy/paste and connect the Image Sampler from the code you posted above into the place I reserved for it.

"Pro Tip": Always work at low-resolution until your algorithms are proven before cranking up to 10K+ geometry counts!

Attached file has low resolution settings with 'Project' and 'SrfSplit' (red group) disabled.

This version (ImageCircles_2017Jun1b.gh) is a dressed up version of the earlier post today, with a small change to the blue group labeled "Remap Sampler".  It now has two sliders instead of one, allowing you to set "Min % Grid" as well as "Max % Grid".  As shown, the smallest circles' radius will be 25% (0.25) of the maximum radius (1/2 of "cell size"), the largest circles' radius will be 90% (0.9) of maximum.