Generating a textual description of an image includes classifying an image represented by image data into a domain-specific category, and segmenting one or more elements in the image data based on the domain-specific category. Each element of the one or more elements is compared to a domain-independent model to detect one or more statistical anomalies in the one or more elements. The one or more detected statistical anomalies are characterized using one or more domain-independent text phrases. The one or more domain-independent text phrases are converted to one or more domain-specific descriptions based upon the domain-specific category.