I was also going to point that out. Any graphics program can blur and image with very similar results.
I could see a benefit to this for pattern recognition, such as determining people's ancestral makeup or what breeds a particular dog is composed of.
The key would be well defined inputs. A large sample of each possible output value would be needed, along with details about a particular value. This would be the training (200 Labradors, 200 Beagles, etc.).
But the next step, testing/usage, requires different software (as far as I know), but their algorithms could probably be re-purposed for it. It would take the new input and compare it against a library of averaged specific samples, somehow determining a % match for each. Then it could provide an estimate of a multi-source makeup (mixed dogs, mixed humans, etc.).
Something of this nature, might be able to identify mixed genetics visually, or to help identify minute genetic differences in a given population (without a ton of tedious study/measurements). I'm thinking of the finches in the Galapagos Islands, that sort of thing. The "purity" of the inputs would be critical. Multiple angles would probably be necessary for animals.
Easy to see negative eugenics type uses (and I think useful applications would be of limited value, maybe).