The purpose of image saliency detection is to obtain high-quality saliency maps that can reflect the significance degrees of different image areas. Based on the saliency map

the visually salient regions of the input images can be processed efficiently

which benefits various applications

such as image segmentation

object detection

and object recognition. According to the theoretical analysis of regional covariance

the intrinsic properties of the image superpixels can be described by the high-dimensional covariance matrix

and thus

the dissimilarity degree between two image superpixels can be determined by the regional covariance distance. Using the regional covariance analysis

a novel method for image saliency detection is proposed. First

the input image is preprocessed by superpixel segmentation. Then

the saliency of superpixels can be calculated using the regional covariance distance. Finally

the saliency of superpixels can be up-sampled to determine the saliency of the image pixels. In this study

we test 200 images selected from the THUS10000 data set for saliency analysis and compare 4 different detection schemes. Experimental results show that our saliency maps are similar to the ground truth manual calibration results. Our method can effectively estimate the saliency of input images with complex background or with similar color between front and background. By combining the high-dimensional intrinsic properties of image pixels and superpixels

our approach can not only avoid the negative effect of single noise pixels but also improve the accuracy of saliency detection. Moreover

by using the covariance matrix of image superpixels

the final saliency map can be robust to the number of feature points

sequence of image pixels

and illumination. The regional-covariance-based image saliency map can be applied to salient object extraction and image segmentation.