Visual saliency detection using iterative outlier cluster elimination

摘要

In saliency region detection, a contrast feature has been widely employed as a standard feature; however, many technical challenges remain when there is high variance between the pixel-inside properties of objects. Recently, an effective hard segmentation (HS)-wise saliency model has been proposed to overcome the drawbacks of using a contrast feature. Although the HS model is quite solid, its heuristic optimization and computationally intensive processes still present limitations. Based on the observations, we propose an iterative outlier cluster elimination for the HS wise saliency computation. The proposed model can be decomposed into the following four phases: regional feature extraction, region clustering, saliency computation, and iterative processing. The motivation of the proposed model is that only good-quality saliency maps generated from the reliable clusters are utilized to optimize the final saliency map by eliminating the outlier clusters. Experimental results demonstrate that the proposed model outperforms state-of-the-art models on various benchmark datasets that comprise images, including with single-, multiple-, and complex-objects. The proposed model is also less computationally complex than existing HS models with minimum performance loss.