Suppress and Balance: A Simple Gated Network for Salient Object Detection

Most salient object detection approaches use U-Net or feature pyramidnetworks (FPN) as their basic structures. These methods ignore two key problemswhen the encoder exchanges information with the decoder: one is the lack ofinterference control between them, the other is without considering thedisparity of the contributions of different encoder blocks. In this work, wepropose a simple gated network (GateNet) to solve both issues at once. With thehelp of multilevel gate units, the valuable context information from theencoder can be optimally transmitted to the decoder. We design a novel gateddual branch structure to build the cooperation among different levels offeatures and improve the discriminability of the whole network. Through thedual branch design, more details of the saliency map can be further restored.In addition, we adopt the atrous spatial pyramid pooling based on the proposed"Fold" operation (Fold-ASPP) to accurately localize salient objects of variousscales. Extensive experiments on five challenging datasets demonstrate that theproposed model performs favorably against most state-of-the-art methods underdifferent evaluation metrics.