A Semantic Segmentation Framework for Small Objects Segmentation in Remote Sensing Images

High spatial resolution (HSR) remote sensing imagery poses challenges for semantic segmentation, particularly concerning background noise reduction, foreground clarity enhancement, and precise identification of densely distributed small objects. To address these issues, we introduce a novel semantic segmentation framework. We use PointFlow modules and the Foreground-Scene (F-S) module to mitigate background interference and enhance foreground saliency. For the challenge of identifying small object clusters, we employ a dual-branch structure: Foreground Precedence Estimation (FPE) and Small Object Edge Alignment (SOEA). Our adaptive loss function, guided by foreground saliency, refines the network's ability to navigate foreground-background segmentation intricacies. Our method surpasses generic segmentation techniques and state-of-the-art remote sensing methods, demonstrating superior performance on the iSAID and Vaihingen datasets.