Saliency detection in images using graph-based rarity, spatial compactness and background prior

Bottom-up saliency detection techniques extract salient regions in an image while free-viewing the image. We have approached the problem with three different low-level cues- graph based rarity, spatial compactness and background prior. First, the image is broken into similar colored patches, called superpixels. To measure rarity we represent the image as a graph with superpixels as node and exponential color difference as the edge weights between the nodes. Eigenvectors of the Laplacian of the graph are then used, similar to spectral clustering (Ng et al., 2001). Each superpixel is associated with a descriptor formed from these eigenvectors and rarity or uniqueness of the superpixels are found using these descriptors. Spatial compactness is computed by combining disparity in color and spatial distance between superpixels. Concept of background prior is implemented by finding the weighted Mahalanobis distance of the superpixels from the statistically modeled mean background color. These cues in combination gives the proposed saliency map. Experimental results demonstrate that our method outperforms many of the recent state-of-the-art methods both in terms of accuracy and speed.