KCP-Net: A novel network Incorporating synergistic attention refinement and progressive adaptive multiscale aggregation for object detection in remote sensing imagery

• The DRF module provides adaptive weights according to different scales. • The KA-SRD combines hierarchical refinement with decomposed large kernel attention. • The CA-DRD uses multi-path fusion and coordinate attention. • The PAA-FPN to aggregates cross-scale features from shallow to deep. Object detection in remote sensing images (RSIs) faces severe scale variations across diverse objects. To address this issue, this paper proposes a novel synergistic attention refinement and progressive adaptive multiscale aggregation network (KCP-Net). KCP-Net consists of three key components. First, kernel-attentive shallow refinement distiller (KA-SRD) is designed to preserve fine-grained details via multi-branch refinement and enlarge receptive field through kernel-attentive driven spatial weighting. Second, coord-attentive deep refinement distiller (CA-DRD) refines deep feature representations via multi-path fusion and coordinate attention to encode directional spatial dependencies and preserve global contextual integrity. Third, a progressive adaptive aggregation feature pyramid network (PAA-FPN) is designed to enhance multiscale feature fusion by integrating hierarchical feature interaction and dynamic-adaptive feature weighting. Extensive experiments on the DIOR, DOTA, and RSOD datasets demonstrate that the AP50 reaches 78.2%, 65.9%, and 92.4% with three datasets, respectively. Comparison experiments between KCP-Net and leading detection models reveal that the average improvements of AP50 are higher 10.5%, AP75 by 12.5%, APS by 5.6%, APM by 10.0%, and APL by 17.2%, suppressing popular one-stage methods (e.g., ATSS, YOLO series etc.) and advancing two-stage methods (e.g., MS-RCNN, Faster-RCNN etc.). Comparison experiments between PAA-FPN and different FPNs demonstrate that the AP50 reaches 76.6%, which is higher by 7.6% than BiFPN, by 5.2% than AFPN. Ablation experiments demonstrate that KA-SRD, CA-DRD, and PAA-FPN increased AP50 by 5.9%, 4.9%, and 5.3%, respectively. These results demonstrate the superiority of KCP-Net in multiscale remote sensing scenarios.