Image Saliency Detection Based on Region Merging

In this paper, an image saliency detection algorithm based on region merging is proposed. Our goal is to divide the image into the background region and the salient region directly via a region merging process. To achieve this, different merging strategies are employed in different stages. Firstly, the image is over-segmented into superpixels. In the first stage, only similar neighboring superpixels are allowed to be merged together. Then, in the second stage, occlusions and hole-regions produced in the last step are handled. At last, instead of merging salient regions together, non-salient regions are merged into the background region guided by the region saliency analysis. The final saliency result can be obtained via a weighted average of several saliency region proposals obtained during the region merging process. Our algorithm is tested on the two public datasets and compared with other state-of-the-art algorithms. The experimental results show that our algorithm is effective and efficient. Particularly, for the more challenging dataset ECSSD, the accuracy of our algorithm outperforms other region-based state-of-the-art algorithms.