Mixing low-level and semantic features for image interpretation