摘要：Saliency detection based on deep learning has achieved excellent performance. This paper focuses on channel-wise feature response. Our model owns three structures: channel-wise extraction (CCFE), channel-wise hierarchical refinement (CHFR), and feature maps fusion (HFMF). The squeeze and excitation residual net is the base network used in this work, which can explicitly explore the dependencies among channels. CCFE is utilized to obtain global feature maps, which has a lot of information loss. Then, the CHFR structure is carried out to take advantage of spatial cues. At the end, HFMF is implemented to fuse the feature maps. The experimental results prove that the proposed algorithm has high efficiency and superior performance compared with other fourteen methods.
