Quantifying Character Similarity with Vision Transformers