Region-level detailed image captioning