Hm. Are NNs reliable enough to be sure you're getting a correct result?

I'd imagine this kind of task would lend itself to some old-school image recognition techniques - photograph against uniform background, threshold, mask the middle part and count contiguous regions.