Abstract
Remote sensing object detection is severely hindered by background clutter and uneven object spatial distribution, limiting the performance of traditional algorithms and the original RT-DETR. To address these issues, this paper proposes an improved RT-DETR-based algorithm, CB-DETR. First, a symmetry-guided Density-Adaptive Attention (DAA) module is designed to tackle insufficient intra-scale feature interaction and poor adaptability to uneven density regions in RT-DETR. Centered on a density estimation network, it predicts target density, generates normalized weights via temperature scaling and softmax, and dynamically adjusts receptive fields through a multi-branch structure to symmetrically adapt to high- and low-density regions, outperforming RT-DETR’s fixed receptive field design. Second, a cross-attention-fused Posterior Dynamic Query Decoder (PDQD) is constructed to overcome fixed query interaction and weak small/occluded object detection in the original decoder. A dynamic query update mechanism optimizes vectors via multi-round iterations, breaking fixed-layer limitations and mining detailed features in complex scenarios, thus improving small/occluded target detection accuracy. Comparative experiments on RSOD, DIOR, and DOTA datasets show that CB-DETR outperforms the original RT-DETR comprehensively: / improve by 2.8%/2.1% and Precision (P)/Recall (R) by 4%/2.4% on RSOD; improves by 1.3% on DIOR and 3% on DOTA. All core metrics surpass the original model and mainstream improved algorithms, verifying the effectiveness and innovation of the proposed improvements.