Text this: Image Captioning Based on Semantic Scenes