Context-based image semantic similarity