Abstract

Traditional1 salient object detection models often use  hand-crafted features to formulate contrast and various  prior knowledge, and then combine them artificially. In this  work, we propose a novel end-to-end deep hierarchical  saliency network (DHSNet) based on convolutional neural  networks for detecting salient objects. DHSNet first makes  a coarse global prediction by automatically learning  various global structured saliency cues, including global  contrast, objectness, compactness, and their optimal  combination. Then a novel hierarchical recurrent  convolutional neural network (HRCNN) is adopted to  further hierarchically and progressively refine the details  of saliency maps step by step via integrating local context  information. The whole architecture works in a global to  local and coarse to fine manner. DHSNet is directly trained  using whole images and corresponding ground truth  saliency masks. When testing, saliency maps can be  generated by directly and efficiently feedforwarding testing  images through the network, without relying on any other  techniques. Evaluations on four benchmark datasets and  comparisons with other 11 state-of-the-art algorithms  demonstrate that DHSNet not only shows its significant  superiority in terms of performance, but also achieves a  real-time speed of 23 FPS on modern GPUs