| 摘 要: 针对航空遥感图像中小目标密集、旋转多变及背景复杂等问题,提出了一种基于改进ConvNeXt的旋转目标检测框架ICRD。使用全维动态卷积(ODConv)在通道、滤波器与空间3个维度自适应分配权重,增强对旋转目标和小目标的特征捕获能力。结合 ConvNeXt的大核深度卷积与 C2f结构的跨层交互特性构建 C2fX模块,以扩展感受野并提升多尺度特征融合效率。在浅层与深层引入全局注意力机制(GAMAttention),以抑制背景干扰并突出目标响应。设计解耦式旋转检测头 CBP-OBB,通过概率交并比(ProbIoU)与中心偏移惩罚(CBP)相结合,优化密集场景下的旋转框定位。实验结果表明:该方法在 DOTA 与 FAIR1M 数据集上平均精度均值(mAP)分别达到77.23%与50.35%,推理速度达到了84frame/s,在准确性和鲁棒性方面超越现有算法。 |
| 关键词: 旋转目标检测 ConvNeXt 全维动态卷积 多尺度特征融合 中心偏移惩罚 |
|
中图分类号: TP391
文献标识码: A
|
|
| Rotated Object Detection in Aerial Imagery via Improved ConvNeXt |
|
WANG Shidie, QI Yunsong, DAI Tao
|
(School of Computer, Jiangsu University of Science and Technology, Zhenjiang 212100, China)
yolanda.wangsd@foxmail.com; mailqys@163.com; leondav1s@163.com
|
| Abstract: To address the challenges of dense small objects, diverse rotations, and complex backgrounds in aerial remote sensing images, this paper proposes ICRD—a rotating object detection framework based on improved ConvNeXt. The framework introduces Omn-i dimensional Dynamic Convolution (ODConv) to adaptively assign weights across channel, filter, and spatial dimensions, thereby enhancing feature extraction capability for rotated and small objects. By integrating the large-kernel depthwise convolution from ConvNeXt and the cross-layer interaction characteristics of the C2f structure, a C2fX module is constructed to expand the receptive field and improve mult-i scale feature fusion efficiency. A Global Attention Mechanism (GAMAttention) is incorporated at both shallow and deep levels to suppress background interference and highlight target responses. Furthermore, a decoupled rotating detection
head named CBP-OBB is designed, which combines Probabilistic Intersection over Union (ProbIoU) with a Center Bias Penalty (CBP) to optimize the localization of rotated bounding boxes in dense scenes. Experiments demonstrate that the proposed method achieves mean Average Precision (mAP )scores of 77.23% on the DOTA dataset and 50.35% on the FAIR1M dataset, with an inference speed of 84 frame/s, outperforming existing algorithms in terms of both accuracy and robustness |
| Keywords: rotated object detection ConvNeXt omn-i dimensional dynamic convolution mult-i scale feature fusion center bias penalty |
