Demystify Self-Attention in Vision Transformers from a Semantic Perspective: Analysis and Application