Learning Generalized Spatial-Temporal Deep Feature Representation for No-Reference Video Quality Assessment