Best of Both Worlds: Learning Arbitrary-scale Blind Super-Resolution via Dual Degradation Representations and Cycle-Consistency