作者
Guoqing Zhou,Longjie Zhu,Xiao Zhou,Yanling Lu,Jón Atli Benediktsson
摘要
Vehicle detection in high-resolution optical remote sensing images faces three challenges: diverse vehicle scales, sparse and scattered distribution, and highly similar backgrounds. To cope with these challenges, this paper proposes CAGR-NET, a gated multi-scale enhancement network. First, a Remote Sensing-Adaptive Modulated Gated Multi-Scale Convolution (RAMGMSConv) combines parallel multi-scale depthwise convolutions with a convolutional gating mechanism to extract multi-granularity features and mitigate the impact of extreme scale variations. Second, a Cross-Association Global Relational Attention (CAGRAttention) module, built on the non-local paradigm with a compact affinity design, models long-range semantic associations to enhance the representation of sparse targets. Finally, an RVAIoU loss adds a scale-aware penalty to CIoU-style geometric constraints to alleviate positional drift. CAGR-NET achieves mAP50 scores of 75.3%, 94.6%, and 87.3% on vehicle subsets of the DIOR, SIMD, and DOTA datasets, respectively. On DIOR, it improves mAP50 by 6.8, 3.3, and 3.0 percentage points (pp) over RT-DETR, YOLOv10, and YOLOv12, respectively; on SIMD, by 6.7 and 1.6 pp over RT-DETR and YOLOv10; and on DOTA, by 3.3 and 3.5 pp over YOLOv10 and YOLOv12. Ablation studies on DIOR show that RAMGMSConv, CAGRAttention, and RVAIoU individually improve mAP50 by 2.1, 2.0, and 1.0 pp, respectively, and the complete framework improves mAP50 by 3.1 pp over the YOLOv11 baseline with 2.622 M parameters. Among the compared methods, CAGR-NET achieves the highest mAP50 and mAP50-95 on all three datasets.