Generating Knowledge-Enriched Image Annotations for Fine-Grained Visual Classification