Spatial Attention Feedback Iteration for Lightweight Salient Object Detection in Optical Remote Sensing Images
Huilan Luo, Jianqin Wang, Bocheng Liang
Abstract
Salient object detection in optical remote sensing images presents distinct challenges, primarily due to the small scale and background similarity of salient objects in images captured by satellite and aerial sensors. Traditional approaches often fail to effectively utilize high-resolution details from shallow features, focusing instead on the semantic depth of features, and typically employ complex, resource-intensive architectures. To overcome these limitations, this article introduces a novel lightweight network, the spatial attention feedback iteration network (SAFINet). SAFINet employs a unique approach by integrating a feature refinement via attention feedback module and a spatial correlation (SCorr) module. The FRAF module refines low-resolution spatial attention (SA) using high-resolution SA, while the SCorr module enhances the fusion of the SAs. These modules work collaboratively to effectively preserve detail integrity and clarity. In addition, a multiscale attention fusion module leverages multiscale information to enrich contextual detail. Our extensive testing on two benchmark datasets shows that SAFINet achieves superior performance in six out of eight metrics, with only 3.12 M parameters and 7.63 G FLOPs, demonstrating significant improvements over 18 state-of-the-art models.