Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation