This paper constructs a new non-uniform Doo-Sabin subdivision scheme via eigen polygon. The authors proved that the limit surface is always convergent and is G1 continuous for any valence and any positive knot intervals under a minor assumption, that λ is the second and third eigenvalues of the subdivision matrix. And then, a million of numerical experiments are tested with randomly selecting positive knot intervals, which verify that our new subdivision scheme satisfies the assumption. However this is not true for the other two existing non-uniform Doo-Sabin schemes in Sederberg, et al. (1998), Huang and Wang (2013). In additional, numerical experiments indicate that the quality of the new limit surface can be improved.