Saliency Detection by Selective Strategy for Salient Object Segmentation

Saliency detection is useful for many computer vision tasks including content-based image retrieval, segmentation, and object detection. However, methods on saliency detection are usually greatly affected by factors like features and segmentation results. We propose a novel selective segmentation-based saliency detection model to decrease the side effects caused by these factors. After extracting different features in LAB color space based on different segmentation results, a series of saliency maps are produced by a segmentation-based method combined with the proposed spatial distributionbased regional saliency measure. The ultimate saliency map is selected from these maps using a novel saliency map evaluation method. The model generates high quality saliency maps that highlight the whole salient object with well-defined boundary. Experiments conducted on Achanta’s dataset show that the model outperforms five state-of-art saliency detection methods on the ground-truth evaluation, yielding better precision-recall curve. We also present application of our saliency maps in an automatic salient object segmentation scheme using Grabcut.