Enhanced salient object detection in remote sensing images via dual-stream semantic interactive network

Salient object detection in remote sensing images (RSI-SOD) aims to identify the most prominent regions within complex RSI scenes. Current convolutional neural network (CNN)-based approaches struggle to capture long-distance dependencies, limiting their performance. To address this, we propose a novel dual-stream semantic interactive network (DSINet). Specifically, the model combines the advantages of Transformer and CNN to simultaneously model both global relationships and local details via the dual-stream architecture. It comprises three key modules: a multi-scale feature enhancement module to enhance feature representations across scales, a cross-attention complementary mining module to explore complementary cues between Transformer and CNN features, and a cross-layer feature interaction module to mitigate inconsistencies between adjacent layers. Extensive experiments on benchmark datasets demonstrate that DSINet achieves superior performance compared to state-of-the-art methods, effectively identifying salient objects in challenging RSI scenes. The code and results of our method are available at https://github.com/dqxfj99/DSINet.