Exploiting Multi-Scale Parallel Self-Attention and Local Variation via Dual-Branch Transformer-CNN Structure for Face Super-Resolution.