Salient object detection network with multi-scale feature refinement and boundary feedback

Abstract Benefiting from the development of convolutional neural networks, salient object detection has achieved prominent progress in recent years. However, there are still two limitations when handling challenging scenarios. On the one hand, upsampling and pooling operations might cause blurry boundaries. On the other hand, multi-level features having different characteristics might incur fusion problems. To address these issues, we propose a novel pixel-wise salient object detection method equipped with multi-scale feature refinement and boundary feedback in this paper. Firstly, a feature interaction scheme is designed to depict the multi-scale representation of side-output features and the interactions between intra-layer features. Secondly, we design a context-aware feature refinement module to adaptively select useful side-output information for generating coarse saliency features. Finally, we obtain the predicted saliency maps with the help of extracted boundary information in a coarse-to-fine manner. Extensive experimental results on four benchmark datasets show that our proposed model performs favorably against state-of-the-art methods.