Ah, I think I see what you mean now. It is probable that most green shades have lower cosine distance amongst themselves than they have with the red pixels of the bird. So using cosine distance might help in finding some of the red colors.
Because there are still drawbacks I would recommend the solution using the L2-distance, but, regardless, what you propose seems reasonable to try to initialize some cluster centers on the bird's pixels.
Because there are still drawbacks I would recommend the solution using the L2-distance, but, regardless, what you propose seems reasonable to try to initialize some cluster centers on the bird's pixels.
