A dual-stream learning framework for weakly supervised salient object detection with multi-strategy integration

Recently, some scribble-based weakly supervised salient object detection (SOD) methods have been proposed to alleviate the heavy burden of expensive and time-consuming pixel-level data labeling in fully supervised SOD. However, due to the lack of salient object structure information in scribble annotations, it is difficult for a model to accurately discriminate and learn explicit boundaries during training. In this paper, we propose a dual-stream learning framework that employs an individual encoding stream to obtain boundary information to help the network identify integral salient regions and accurate structural details. Additionally, we adopt different strategies (i.e., the boundary-aware semantics enhancement module for the high levels, the boundary-aware detail enhancement module for the low levels) to better integrate boundary information with object features at different levels to take full advantage of different salient object feature properties. Extensive experiments show that our model achieves competitive performance against the state-of-the-art weakly supervised SOD methods, demonstrating the superiority and effectiveness of our proposed network. The code and results are released from the link: https://github.com/boom118/BSnet.