Densely Connected Refinement Network for Salient Object Detection
Qinghua Ren, Renjie Hu · 2018
Fully Convolutional Neural Network (FCN) has recently made a great breakthrough in salient object detection. However, modern deep learning based methods usually combine low-level edge information and high-level semantic knowledge in a simple way, leading to the distractions from background in some challenging cases. In this paper, we propose a densely connected refinement network to make full use of deep feature derived from multiple convolutional layers. By adopting the dense connectivity strategy, the semantic information from deeper layers can be directly passed through to shallower layers. These short connections can effectively strengthen the feature propagation during the training process. Moreover, the proposed model introduces fewer parameters to achieve a real-time computation speed while guaranteeing outstanding performance. Quantitative and qualitative experimental results on 4 benchmark datasets demonstrate that our approach compares favorably against other top-performing methods.