TDA-DETR: target-context guided aggregation and dynamic multi-scale feature interaction for remote sensing object detection

Abstract Remote sensing imagery serves as a crucial medium for geophysical observation. However, its inherent complexities, including pronounced scale variations, ambiguous object boundaries, and high visual similarity to complex backgrounds, pose significant challenges to precise object detection. These factors weaken the discriminative capability of object features and hinder the effective utilization of contextual information, particularly when detecting small and sparsely distributed targets. To overcome these challenges, this study proposes a novel remote sensing object detection framework, TDA-DETR, based on RT-DETR. The proposed framework incorporates a guided aggregation and dynamic feature interaction mechanism to enhance both detection accuracy and robustness, which are essential for ensuring reliable measurements in complex scenes. It achieves its objectives through a series of collaboratively designed core modules. First, the target-context guided feature aggregation module introduces a novel dual-guided serial attention paradigm, which leverages Haar wavelet-decoupled signals to deeply integrate target and contextual information. Second, the dynamic multi-scale interaction fusion block, as the core of the DFusion module, efficiently consolidates multi-scale features through parallel transformations and interactive mechanisms. Finally, the adaptive sparse attention-based intra-scale feature interaction module dynamically combines sparse and dense attention to refine features and suppress background noise. Experiments on three public remote sensing datasets—DIOR, NWPU very high resolution-10, and RSOD—demonstrate that TDA-DETR achieves a mean average precision of 77.6%, 95.4%, and 95.7% at an Intersection over Union threshold of 0.5, substantially outperforming baseline models and other state-of-the-art methods.