Detecting salient regions in static images
Guangchun Cheng, Eric Ayeh, Ziming Zhang · 2012
Computational visual attention systems, which aim at detecting salient regions in images, have been the subject of research for more than two decades. In this paper, we propose a novel approach (SEC) to detect salient regions in static images. This method is composed of two modules: segmentation-based entropy computation to determine the information content of clusters and local color contrast computation to enhance the saliency. DBSCAN is used first to segment the image. Then, the entropies of the resulting segments are computed. Spatial information of each segment size and cohesion is employed to adjust the entropy in terms of distinctiveness. Color contrast between adjacent segments is then computed and combined with spatial information to determine the most salient regions within the input image. We conducted two types of experiments, and compared visually and quantitatively with existing methods.