Dense Residual Pyramid Networks for Salient Object Detection

We introduce a coarse-to-fine method for salient object detection. In fully convolutional networks (FCN), pooling operation generates downsampled feature maps, while full size estimation is required for salient objet detection. Our Dense Residual Pyramid Networks (DRPN) attends to generating high-resolution and high-quality results. However, in order to provide enough local information, we extract extra local features from pre-trained networks. Finally, the proposed dense residual blocks learn to merge all the information and generate full size saliency maps. In our work, the thought of reconstructing Gaussian pyramids is first introduced into the frameworks of convolutional neural networks. We employ dense residual learning to learn residual maps. We hope these feature maps can be used to refine the upsampled feature maps, as Laplacian images can be used to reconstruct images in Gaussian pyramids. Experiments show that our DRPN has huge improvement over previous state-of-the-art methods on all the datasets. Especially, our DRPN outperforms previous state-of-the-art over 11.6% on ECSSD.