Learning Hybrid Representations of Semantics and Distortion for Blind Image Quality Assessment