AR-YOLOv11-OBB: A Remote Sensing Oriented Object Detection Method via Adaptive Rectangular Convolution

Objects in remote sensing images are often characterized by arbitrary orientations, significant variations in aspect ratios, and dense spatial distributions, which make conventional horizontal bounding boxes inadequate for accurate spatial representation. Although YOLO-based rotated object detection methods have demonstrated high detection efficiency, their backbone networks typically rely on fixed square convolution kernels, limiting their ability to model the anisotropic geometric characteristics of remote sensing targets. To address this issue, this paper proposes an improved rotated object detection model termed AR-YOLOv11-OBB. Built upon YOLOv11-OBB, the proposed method introduces Adaptive Rectangular Convolution (ARConv) into the backbone network, enabling the height and width of convolution kernels to adapt dynamically to feature content. This design allows better alignment with the spatial structures of elongated and rotated objects. Compared with deformable convolution, ARConv achieves dynamic receptive field shape modeling while maintaining lower computational complexity. Extensive experiments conducted on the DOTA v1.5 dataset demonstrate that the proposed method outperforms multiple mainstream rotated object detection models in terms of detection accuracy, exhibiting particularly stable performance in scenarios involving elongated objects and densely distributed targets.