A Semantic Higher-Level Visual Representation for Object Recognition