Scholarlink

With the rapid development of unmanned aerial vehicle technology, aerial object detection has become crucial across multiple domains. However, challenges such as extreme scale variations, dense small objects, and complex backgrounds severely limit existing detection methods. This paper proposes DRA-YOLO, a lightweight multi-scale feature enhancement framework that pioneers the application of YOLOv11 architecture to aerial object detection. The framework synergistically integrates three core innovations: Dynamic Receptive-field Attention Convolution (DRAConv) for adaptive multi-scale feature extraction, Dynamic Gated Feature Pyramid Network (DyGateFPN) with dedicated small object detection branches and dynamic upsampling for enhanced feature fusion, and Dynamic Receptive-field Detail-Enhanced Convolutional Detection Head (DRDECD) for precise fine-grained localization. Extensive experiments on four aerial benchmarks (VisDrone, AI-TOD, SIMD, and DIOR) demonstrate that DRA-YOLO achieves superior accuracy-efficiency trade-offs with only 3.03M parameters and 13.9 GFLOPs. The source code is available at https://github.com/claudepsis/dra_yolo.git .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DRA-YOLO: Dynamic Receptive-Field Attention and Dual-Gated Upsampling Model for Aerial Object Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qifeng Shi,
  • Yan Zhang

摘要

With the rapid development of unmanned aerial vehicle technology, aerial object detection has become crucial across multiple domains. However, challenges such as extreme scale variations, dense small objects, and complex backgrounds severely limit existing detection methods. This paper proposes DRA-YOLO, a lightweight multi-scale feature enhancement framework that pioneers the application of YOLOv11 architecture to aerial object detection. The framework synergistically integrates three core innovations: Dynamic Receptive-field Attention Convolution (DRAConv) for adaptive multi-scale feature extraction, Dynamic Gated Feature Pyramid Network (DyGateFPN) with dedicated small object detection branches and dynamic upsampling for enhanced feature fusion, and Dynamic Receptive-field Detail-Enhanced Convolutional Detection Head (DRDECD) for precise fine-grained localization. Extensive experiments on four aerial benchmarks (VisDrone, AI-TOD, SIMD, and DIOR) demonstrate that DRA-YOLO achieves superior accuracy-efficiency trade-offs with only 3.03M parameters and 13.9 GFLOPs. The source code is available at https://github.com/claudepsis/dra_yolo.git .