Enhancing semantic features with compositional analysis for scene recognition