Saliency Detection Using Fully Convolutional Network

Abstract

A sophisticated saliency detection method based on a fully convolutional network is proposed. First, an end-to-end network model is trained, by which an initial saliency map of the input image is yielded. Then, the accuracy of object boundaries in the initial saliency map is improved by using the fully connected conditional random field. As a result, an intermediate saliency map with more precise edges is obtained. Finally, a saliency cut technique is exploited to further improve the performance of the saliency map. Extensive experiments conducted on four benchmark image datasets and in the presence of different levels of noise show that the proposed method can perform better than a number of state-of-the-art saliency detection algorithms.