Refining Text-to-Image Generation: Towards Accurate Training-Free Glyph-Enhanced Image Generation