Optimal Feature Selection for Saliency Seed Propagation in Low Contrast Images

Abstract

Listen

Translate article

Salient object detection can substantially facilitate a wide range of applications. Although significant improvements have been made in recent years, low contrast image still pose great challenges to current methods due to its low signal to noise ratio property. In this paper, an optimal feature selection based saliency seed propagation method is presented to detect salient objects in low contrast images. The key idea of the proposed approach is to hierarchically refine the saliency map guided by adaptively selecting the optimal features in low contrast images recursively. Multiscale superpixel segmentation is firstly utilized to suppress background interference. Then, the initial saliency map can be generated via global contrast and spatial relationship. Local and global fitness are finally utilized to optimize the resulting saliency maps. Extensive experimental evaluations on four datasets demonstrate that the proposed model outperforms 15 state-of-the-art methods in terms of efficiency and accuracy.