Scribble-attention hierarchical network for weakly supervised salient object detection in optical remote sensing images

Arising from cluttered background interference and various scaled objects, salient object detection (SOD) in optical remote sensing images (RSIs) is a challenging task. In the latest research, supervision-based methods have made significant progress due to elaborate annotations. However, these supervision-based methods still face the following two problems: 1) The elaborate annotations require a lot of resources to label, and inaccurate labeling can result in serious training error, 2) the annotations of regular shape are not suitable for labeling the object of different scenes. In this paper, a scribble-attention hierarchical network (SHNet) is proposed to tackle the above issues. Firstly, instead of elaborate annotations, we relabel an optical RSI dataset in a scribble manner. Secondly, a novel scribble embedding network (SEN) based on correlation filtering technology for extracting discriminant regions is designed, which can illuminate the foreground regions and suppress the cluttered background regions while increasing the perceptual field of objects. Finally, we present a multi-branch edge detection network (MBED) to enhance the localization of various scaled objects and get sharp edges. By fusing the output features of SEN and MBED, the distinction between the foreground and background of various scaled objects is strengthened. Experimental results on existing optical RSI datasets verified the effectiveness of our proposed SHNet. The source code is publicly available at .