Salient Region Detection Based on Spectral Residual and Sparse Reconstruction Error

In this paper, we propose a salient region detection method based on spectral residual and sparse reconstruction error. Spectral residual is obtained in the frequency domain, and very effective in extracting foreground which is relatively more salient than background in an image. Sparse reconstruction error effectively represents the whole image because of considering sparsity. We combine both spectral residual and sparse reconstruction error to successfully detect salient regions in an image. We first choose salient super-pixels obtained by image segmentation by spectral residual, and then compute sparse reconstruction error based on the salient super-pixels. Finally, we generate visual saliency maps by normalizing the visual saliency. Experimental results demonstrate that the proposed method successfully detects salient regions in an image and outperforms state-of-the-art ones in terms of the average score in subjective evaluations.