Hierarchical Salient Object Detection Network with Dense Connections

Abstract

Fully convolutional neural networks (FCNs) have shown outstanding performance in many dense labeling tasks. FCN-like salient object detection models haven mostly developed lately. In the work, we propose a novel pixel-wise salient object detection network based on FCN by aggregating multi-level feature maps. Our model first makes a coarse prediction by automatically learning various saliency cues, including color and texture contrast, shapes and objectness. Then a densely connected feature extraction block is adopted to further extract rich features at each resolution. Moreover, skip-layer structure is introduced for providing a better feature representation and helping shallow side outputs locate salient objects. In addition, a weighted-fusion module is utilized to combine multi-level features. Finally, a fully connected CRF model can be optimally incorporated to improve spatial coherence and contour localization in the fused saliency map. The whole architecture works in a coarse to fine manner. Evaluations on five benchmark datasets and comparisons with 10 state-of-the-art algorithms demonstrate the robustness and efficiency of our proposed model.