A Fully Convolutional Network based on Spatial Attention for Saliency Object Detection

The performance of the saliency object detection model is suppressed by the false and missed detections in saliency areas. How to effectively reduce the false predictions is an open issue. Unlike existing methods which add hand-craft features or post-processing steps, we propose a novel saliency object detection model based on the fully convolutional network and spatial attention. We first design a branch based on spatial attention which can enable the network to focus on the areas of saliency object. We further design a boundary branch aiming at get accurate object boundaries. Similar to spatial attention branch, we use one-dimensional convolution kernels to capture semantic information in different directions and reduce the number of parameters. We evaluate the model proposed in this paper on four common standard datasets with four common evaluation criteria. Experimental results show that the model we proposed outperforms the state-of-the-arts with less false predictions and accurate object boundaries.