Top-down visual saliency via joint CRF and dictionary learning | IEEE Conference Publication | IEEE Xplore