Semantic Decoupling Guidance Network for Dense Tiny Object Detection

Dense tiny object detection in remote sensing images is fundamentally challenged by low signal-to-noise ratios and cluttered backgrounds, where semantic context is crucial yet difficult to leverage effectively. Existing methods often struggle with feature conflicts when incorporating semantic information. In this work, we propose the Semantic Decoupling Guidance Network (SDGDNet), a novel framework that utilizes semantic priors from a frozen, pre-trained segmentation network for guidance while adaptively integrating the semantic priors. SDGDNet features two core modules: a Mixture-of-Experts Gated Adapter (MEGA) adaptively filters pixel-level semantic features via dynamic routing to bridge the domain gap with instance-level detection, and a Window-based Deformable Attention Guidance (WDAG) module employs the filtered semantics to guide deformable attention and improve robustness against noise and local geometric variations in detection features. Experiments on the DTOD benchmark demonstrate state-of-the-art performance, achieving an AP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> of 31.5%, surpassing previous methods by 7.3% and validating the effectiveness of our semantic decoupling guidance approach.