It propably depends on the geometry you're working with, but you might have a chance in abusing the fact that normal maps are indeed just encoded direction vectors, which are assigned to the R, G and B color slots. If you try a node setup like this (so feeding the normal output into an emission shader and previewing that) on your texture...
... you'll see the famous mostly blue normals texture:
Add two more nodes afterwards to split up the three channels into its components. One is a SeparateRGB node, the other one an invert node. You'll need the invert just on the B channel, not on the other two:
The result of that separation will be this:
Whereas the original texture which I was feeding into the whole chain was simply this:
The background is, in the R channel or a normal map, the perturbation amount of the normal in U direction is stored, whereas in the G channel the V perturbation is recorded. B is - I believe - a factor of how unperturbed the normal is. That's why you need to invert it, as you're interested in the bumpy areas only.
With your provided texture, the setup produces this: