Image Captioning

Image captioning studies how to turn images into useful descriptions, from brief scene summaries to detailed accounts of objects, attributes, relations, and changes. Common approaches combine vision encoders with Transformer or large vision-language model decoders; retrieval, region grounding, and reinforcement learning can add context or encourage more complete, faithful captions. Models are tested on datasets such as MS COCO, Flickr30k, and NoCaps, as well as specialized medical, remote-sensing, and multilingual benchmarks, using both text-overlap metrics and human or question-answering evaluations. A central challenge is balancing detail and coverage against hallucination, while evaluation remains difficult because reference captions are incomplete and automatic metrics can miss visual errors.