Estimation method of initial labels for propagation-based saliency detection
Yo Umeki, Taichi Yoshida, Masahiro Iwahashi · 2016
We propose an estimation method of initial labels based on scale-invariant feature transform (SIFT), high dimensional color transform (HDCT), and machine learning for propagation-based saliency detection. The label propagation strategy is efficient for saliency detection, but its accuracy depends on the distribution of initial labels. In this paper, the proposed method respectively estimates initial labels of fore/background based on the machine learning with HDCT and the density of SIFT feature points. Consequently, initial labels are certainly distributed all over the region and inaccurate saliencies are attenuated in resultant saliency maps. Through simulations, we show that the proposed method outperforms the state-of-the-art methods objectively and perceptually.