Improved salient object detection based on background priors

Abstract

Recently, many saliency detection models use image boundary as an effective prior of image background for saliency extraction. However, these models may fail when the salient object is overlapped with the boundary. In this paper, we propose a novel saliency detection model by computing the contrast between superpixels with background priors and introducing a refinement method to address the problem in existing studies. Firstly, the SLIC (Simple Linear Iterative Clustering) method is used to segment the input image into superpixels. Then, the feature difference is calculated between superpixels based on the color histogram. The initial saliency value of each superpixel is computed as the sum of feature differences between this superpixel and other ones in image boundary. Finally, a saliency map refinement method is used to reassign the saliency value of each image pixel to obtain the final saliency map for images. Compared with other state-of-the-art saliency detection methods, the proposed saliency detection method can provide better saliency prediction results for images by the measure from precision, recall and F-measure on two widely used datasets.