Salient object detection based on multi-scale super-pixels and saliency propagation

A novel method to detect the most salient object from an image is proposed. Saliency cues such as spatial colour distribution and the saliency prior are extended to super-pixels and are integrated to compute the final saliency. The super-pixels are generated in multiple scales so that the proposed method is adaptive to objects of different sizes. Moreover, to solve the problem that many saliency detection algorithms rely on local evidences, it is proposed to propagate the saliency information based on the spectral affinities of super-pixels. Experimental results have proven that the proposed method achieves discriminative saliency maps and outperforms the state-of-the-art methods.