Multi-level annotation of natural scenes using dominant image components and semantic concepts