Heatmap and edge guidance network for salient object detection

Most of existing salient object detection methods are based on convolutional neural networks, which have attained good results. However, most of them suffer from coarse object boundaries, and even predict salient objects as background. In this paper, we propose a novel network named HENet to better extract and utilize the features of different layers to achieve better prediction results. In order to better use features of different levels, we propose a feature extraction module to use heatmap and edge feature as intermediate supervision to get location and detailed information. Then we propose a multi-layer feature supplementary module to add and merge above information to each layer to strengthen the corresponding feature learning ability. What's more, we put forward the trisection dilated convolution module to deal with features to expand the receptive field of features in each layer. We have tested 8 methods on 4 datasets, the experimental results of our method also demonstrate superiority.