Salient object detection using recursive regional feature clustering

In the past decade, contrast features, which focus on rarity and uniqueness, have been widely used in the saliency-detection field, but the extreme dependency on the most-highlighted region remains as a limitation for the detection of multiple and complex objects. In particular, the difficulties are commonly observed when a high inter-object dissimilarity exists. Based on this observation, we present a new paradigm for the detection of the salient-object region, whereby only the spatial-saliency clues are interpreted from a multiple-level clustering framework; for this reason, in contrast to the previous methods, the proposed model is not dependent upon the contrast features. The proposed model can be roughly decomposed into the following four main phases: regional feature extraction, homogeneous-region clustering, saliency-score computation, and recursive processing. In particular, a recursive processing for which the salient region is optimized through the improvement of its clustering results is introduced. According to the experiment results, the proposed scheme outperforms the state-of-the-art methods on various benchmark datasets consisting of single, multiple, and complex object images; furthermore, the proposed model is more effective for the detection of multiple objects. To validate the contributions of this study, a multiple salient-object dataset (MSOD) that contains 100 images with more than two objects with a higher dissimilarity was also constructed.