# Inner working behind combining two distance function as one function for similarity measure

I am comparing two images and for this I am testing various similarity function. For my case, Euclidean works much better than cosine(20% difference). However, I tried to combine two distance function namely- Euclidean+cosine and I found that this new similarity measure performs better than euclidean and cosine. What could be the reason that cosine performs so badly, and combining Euclidean+cosine outperforms other metrics? Is there some mathematical or logical explanation for this? Note: I am comparing one clean image with a noisy image .