Consensus exploration and detail perception for co-salient object detection in optical remote sensing images

Abstract

Co-salient object detection (CoSOD) in optical remote sensing images (ORSI) aims to identify common salient objects across a set of related images. To address this, we introduce the first large-scale dataset, CoORSI, comprising 7668 high-quality images annotated with target masks, covering various macroscopic geographic scenes and man-made targets. Furthermore, we propose a novel network, Consensus Exploration and Detail Perception Network (CEDPNet), specifically designed for CoSOD in ORSI. CEDPNet incorporates a Collaborative Object Search Module (COSM) to integrate high-level features and explore collaborative objects, and a Feature Sensing Module (FSM) to enhance salient target perception through difference contrast enhancement and multi-scale detail boosting. By continuously fusing high-level semantic information with low-level detailed features, CEDPNet achieves accurate co-salient object detection. Extensive experiments demonstrate that CEDPNet significantly outperforms state-of-the-art methods on six evaluation metrics, underscoring its effectiveness for CoSOD in ORSI. The CoORSI dataset, model, and results will be publicly available at https://github.com/chen000701/CEDPNet.