One-stop multiscale reconciliation attention network with scribble supervision for salient object detection in optical remote sensing images

Salient object detection in optical remote sensing images (RSI-SOD) faces significant challenges due to the unique characteristics of RSI imaging. Existing methods heavily rely on labor-intensive pixel-level annotations and overlook the potential of low-cost sparse annotations. Moreover, weakly supervised RSI-SOD methods introduce multiple sparse annotations and training processes, leading to a multistaged SOD task and considerable performance gaps compared to fully supervised approaches. To address these issues, we propose a one-stop end-to-end RSI-SOD method that solely relies on scribble annotations. Our framework, named the one-stop multiscale reconciliation attention network (OMRA-Net), features encoding, reconciliation, polishing, and convergence layers for effective feature extraction, reconciliation, polishing, and object structure restoration. Evaluation on publicly available datasets demonstrates that OMRA-Net outperforms existing weakly supervised and unsupervised SOD methods, achieving comparable or superior performance to fully supervised models. Ablation studies further validate the effectiveness of our proposed model design.