Edge semantic collaboration network for salient object detection in optical remote sensing images

Abstract

The rapid development of deep learning has promoted the development of salient object detection in optical remote sensing images (ORSI-SOD). However, ORSI-SOD faces many challenges, including the interference of color and shadow backgrounds, or the uncertainty of the number and scale of objects in optical remote sensing images (ORSIs). Most of the existing models have difficulty in establishing effective long-distance feature dependencies. To address this issue, we propose an edge semantic collaboration network (ESCNet). Specifically, ESCNet designs an Interactive Graph Inference Module (IGIM) to model channel interactions and capture long-distance semantic dependencies via graph inference. Then, a Semantic Feature Enhancement Module (SFEM) is adopted to refine the dependency information based on a composite attention mechanism. Simultaneously, a Multi-scale Edge Refinement Module (MERM) extracts precise boundaries using multi-scale feature refinement. Finally, the features produced at each stage are sequentially fed into the decoder and generate the final saliency maps. Extensive experiments on three public datasets (ORSSD, EORSSD, and ORSI-4199) confirm the superiority of the proposed ESCNet compared with state-of-the-art methods.