Multicriteria saliency detection: a (exact) robust network design approach

In the last decades, a wide body of literature has been devoted to study different saliency detection methods. These methods are typically devised on the basis of different image analysis paradigms, which leads to different performances that are not always rankable, but rather complementary. In this paper, a network design-based framework for multicriteria robust saliency detection is proposed. The key idea is that a suitable blending of the salient regions obtained by different methods leads to a salient region that outperforms the results obtained, individually, by these methods. Moreover, besides of considering state-of-the-art saliency detection approaches, a new method, which incorporates a novel tool for image contour detection, is designed. Results obtained on different sets of benchmark instances show that the proposed multicriteria robust framework exhibits high accuracy in the detection of salience objects; i.e., the pixels comprising the blended salient object are likely to be part of the actual salient object. This work aims at building further bridges between the areas of image processing and the areas of operations research.