Saliency Detection via Low-rank Reconstruction from Global to Local

Saliency detection can be a useful technique for image semantic analysis such as auto image segmentation, image resize, advertising design and image compression. It is a core problem of saliency computing how to obtain the effective salient object with less non-saliency information, which is consist with movement of eye fixation. In this paper, we propose a saliency computing model based on rank-sparsity decomposition. In order to highlight saliency objects, the model eliminates the non-saliency background information from global to local. In an image, the salient object often has more strong contrast or difference relative to in the background. Firstly, with simple contrast in CIELab color space, we can obtain preliminary map. Secondly, using Low-rank reconstruction in global image, positioned roughly salient object. Finally, in order to eliminate nonsignificant noise, the mode reconstructs the redundant background from the block in the image. The experimental result shows that the proposed method can get a better saliency map compared with the-state-of-arts.