Exploiting Randomized Prim’s Algorithm and Background Contrast for Saliency Detection

ABSTRACT

The geodesic saliency method in the literature was based on the boundary and connectivity priority, which assumed that most of the background regions can touch the image boundaries. It cannot deal with the images with complex backgrounds or variant textures. To address such problem, we propose an improved saliency detection method by involving the important foreground priority. First, the statistical results of randomized Prim’s algorithm are used to generate a coarse conspicuity map, which aims to roughly estimate the potential foreground. Then, the image is over-segmented into some individual superpixels and an affinity propagation clustering method is used to group the superpixels having a similar color appearance together. This is followed by the foreground probability map computation through the spatial interaction information between the coarse conspicuity map and superpixel based color clusters. The final saliency map is generated by integrating the above foreground probability map and background color contrast in a unified way. The quantitative and qualitative comparisons on the benchmark dataset MSRA-1000 and SED show that our method outperforms many recent proposed state-of-the-art approaches significantly.

Keywords:

Background contrast, coarse conspicuity map, foreground probability map, randomized prim’s algorithm, saliency detection.