Frequency and Edge Guided Network for Weakly Supervised Salient Object Detection

Recently, due to the large amount of manpower and material resources required for fully supervised salient object detection in pixel-level label labeling, some weakly supervised salient object detection (SOD) based on scribble annotations have been proposed to reduce the burden. However, since scribble annotations are sparse annotations, they lack salient object structure information, making it difficult for the model to accurately distinguish the foreground from the background, especially on the object boundary. This paper proposes a boundary and frequency domain guided learning framework to obtain information to help the network locate the object, improve the recognition rate, refine the edge and strengthen the capture of structural details. In addition, the edge and frequency domain are fused to make up for the lack of lowlevel features and make full use of the dominant attributes of different modes. Extensive experiments show that our model achieves competitive performance against the state-of-the-art weakly supervised SOD methods, demonstrating the superiority of effectiveness of our proposed network.