GGRNet: Global Graph Reasoning Network for Salient Object Detection in Optical Remote Sensing Images

The task of salient object detection (SOD) in optical remote sensing images (RSIs) is more challenging than the SOD in natural sensing images (NSIs) because of the unique characteristics of remote sensing images such as various object scales and background context redundancy. However, the existing methods ignore the global relationship modeling between different salient objects or different parts in one salient object. To this end, we design a Global Graph Reasoning Module (GGRM) in a lightweight and effective form, and propose a novel Global Graph Reasoning Network (GGRNet) for SOD in optical RSIs. During the graph reasoning, the GGRM considers the role of the global information. Specifically, we explore two ways to utilize the global information, including the global features and global nodes, which are ingeniously added to the interaction of graph nodes and fully integrated through iteration. Besides, we stabilize the projection channel between coordinate space and interactive space through an attention mechanism. The GGRNet outperforms the existing state-of-the-art SOD algorithms on two publicly available datasets, and the number of parameters is only 25.01 Mb.