image-to-text

Image-to-text extracts or generates text from images through OCR or visual captioning/description. **Two meanings**: **OCR**: Extract printed/handwritten text from documents, signs, screenshots (text literally in image). **Captioning**: Generate natural language descriptions of visual content (what the image shows). **OCR technology**: Deep learning OCR (Tesseract, EasyOCR, PaddleOCR), document AI (AWS Textract, Google Document AI), scene text recognition. **Captioning models**: BLIP, BLIP-2, LLaVA, GPT-4V, Gemini Vision - vision-language models generating descriptions. **Dense captioning**: Describe multiple regions of image in detail. **Visual QA**: Answer specific questions about image content. **Document understanding**: Extract structured information from forms, tables, invoices. **Implementation**: Vision encoder + language decoder, cross-attention or prefix tuning, trained on image-caption pairs. **Use cases**: Accessibility (alt-text), content moderation, visual search, document digitization, photo organization. **Evaluation metrics**: BLEU, CIDEr, SPICE for captioning. **Challenges**: Hallucination in descriptions, fine-grained details, counting accuracy. Foundation for multimodal AI applications.