Saliency detection by tensor voting based Gaussian modeling
Huynh Trung Manh, Jungyeon Yeo, Guee-Sang Lee · 2015
Salient object detection is a challenging task in field of object segmentation and recognition. Many works have been proposed but remaining serious limitations. The answers for question:" How to highlight the entire salient object which has captured human eye?" is still unseen until now. In this paper, therefore, we propose a novel, efficient and simple saliency detection method using Gaussian Mixture Modeling based on tensor voting. At first, the color images are mapped into 2-D space in which tensor voting process is applied to find out extremes. Then, Gaussian Mixture Model (GMM) is estimated and for each pixel, the set of normalized likelihood measures to different Gaussian Models are calculated. The color saliency value measure and spatial saliency measure of each Gaussian model are evaluated based on its color distinctiveness and spatial distribution. Finally, the final saliency map is generated by fusing color saliency map and spatial saliency map. We compare our algorithm to several states-of- the art saliency detection methods using the well-known 1000 available public images. The experimental results show that our method significantly outperforms the previous algorithm in both precision and recall. We also show the advantages of our work on object segmentation.