计算机科学
卷积(计算机科学)
特征(语言学)
遥感
目标检测
对象(语法)
干扰(通信)
代表(政治)
人工智能
可视化
计算机视觉
特征提取
钥匙(锁)
遥感应用
感知
可分离空间
卷积神经网络
计算复杂性理论
人工神经网络
作者
Yaru Zhi,Aimin Li,Dongyu Zhuang,Haifeng Wang
标识
DOI:10.1109/lgrs.2026.3754068
摘要
Remote sensing object detection faces several challenges, such as large variations in object scales, insufficient fine-grained feature representation for small objects, and interference from complex backgrounds. To tackle these problems, this paper proposes DSDP-Net, a dynamic scale-detail perception network for lightweight remote sensing object detection. DSDP-Net consists of three main modules. Firstly, the Depthwise Separable Omni-Dimensional Dynamic Convolution C2f module (DS-ODConvC2f) is designed to strengthen multi-scale feature modeling while reducing model parameters and computational cost. Then, the Spatial Fine-Grained Enhancement module (SFGE) is constructed to enhance key regional features and local fine-grained details. Finally, the Dual-Path Ghost-Shuffle Convolution module (DP-GSConv) is developed to realize lightweight dual-path feature interaction and reduce computational redundancy. Experimental results on the NWPU VHR-10, VEDAI, and DOTA-v1.0 datasets demonstrate the effectiveness of DSDP-Net in complex remote sensing scenarios. The network achieves 92.9% mAP50on NWPU VHR-10, 58.6% mAP50on VEDAI, and 67.90% mAP50on DOTA-v1.0, with only 2.76M parameters and 7.6 GFLOPs. Class-wise, scale-wise, ablation, and visualization results further verify the effectiveness of the proposed modules.
科研通智能强力驱动
Strongly Powered by AbleSci AI