SimAN: Exploring Self-Supervised Representation Learning of Scene Text via Similarity-Aware Normalization.