A Fully Convolutional Network based on Double Attention for Saliency Object Detection
Yufeng Wu, Yunfeng Nie, Jianlu Fu, Wenxuan Gong · 2021
The current salient object detection method is based on full convolution network, and the method based on full convolution directly applies multi-layer convolution features without distinguishing, which leads to the result is not optimal because of the dispersion of redundant details; In this paper, based on the existing full convolution network, we first design attention module which can generate attention features by combining channel attention and spatial attention. We further design a boundary branch aiming at get accurate object boundaries. We selectively superimpose information from multi-layer features, optimize attention features by using low-level features, and fuse multi-scale features; This method is compared with six related methods on four common datasets. Experimental results show that the model we proposed outperforms the state-of-the-arts with less false predictions and accurate object boundaries.