Existing solutions like Tesseract already can embed text into the image, but I'm wondering if there's a way to combine LLM with Tesseract, so that LLMs can help correcting results and finding unidentified text, and finally still embed text back to the image