Scholarlink

<p>Object detection in aerial imagery remains challenging due to the dominance of small objects, large scale variations, and the loss of spatial detail during deep feature extraction. While recent Detection Transformer (DETR) architectures have advanced end-to-end detection, they often suffer from weak spatial retention and limited multi-scale fusion in aerial scenes. To address these issues, we propose the Enhanced Hybrid Attention-based DETR (EHA-DETR). First, we design an Enhanced PConv–LFEM Network (EPLN) that integrates pinwheel-shaped convolution (PConv) and a Local Feature Enhancement Module (LFEM) to preserve directional and fine-grained cues, reducing the loss of features for small objects in the backbone. Then, we propose an Adaptive Multi-Scale Fusion Network (AMSFN) that employs learnable bidirectional fusion and spatial-aware weighting, integrating hierarchical features across scales with multi-kernel convolutions to enhance representations of small objects. Finally, we develop a Cross-Scale Feature Enhancement Module (CSFEM) that refines high-resolution features with multi-grained attention guided by high-level semantic features to improve representations of small objects. Experiments on three aerial benchmarks—VisDrone2019, CODrone, and SIMD—validate the effectiveness of EHA-DETR. On VisDrone2019, it achieves 40.6% mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> </InlineEquation>, outperforming the Real-Time Detection Transformer (RT-DETR) baseline by 4.1% while reducing parameters by 16.6%. Consistent improvements on CODrone and SIMD further demonstrate its robust generalization and efficient performance for aerial vision tasks.</p>

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

Eha-detr: an enhanced hybrid attention transformer for small object detection in aerial images

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Xuexiang Li,
  • Xin Li,
  • Xianfu Chen,
  • Pengsong Duan

摘要

Object detection in aerial imagery remains challenging due to the dominance of small objects, large scale variations, and the loss of spatial detail during deep feature extraction. While recent Detection Transformer (DETR) architectures have advanced end-to-end detection, they often suffer from weak spatial retention and limited multi-scale fusion in aerial scenes. To address these issues, we propose the Enhanced Hybrid Attention-based DETR (EHA-DETR). First, we design an Enhanced PConv–LFEM Network (EPLN) that integrates pinwheel-shaped convolution (PConv) and a Local Feature Enhancement Module (LFEM) to preserve directional and fine-grained cues, reducing the loss of features for small objects in the backbone. Then, we propose an Adaptive Multi-Scale Fusion Network (AMSFN) that employs learnable bidirectional fusion and spatial-aware weighting, integrating hierarchical features across scales with multi-kernel convolutions to enhance representations of small objects. Finally, we develop a Cross-Scale Feature Enhancement Module (CSFEM) that refines high-resolution features with multi-grained attention guided by high-level semantic features to improve representations of small objects. Experiments on three aerial benchmarks—VisDrone2019, CODrone, and SIMD—validate the effectiveness of EHA-DETR. On VisDrone2019, it achieves 40.6% mAP \(_{50}\) , outperforming the Real-Time Detection Transformer (RT-DETR) baseline by 4.1% while reducing parameters by 16.6%. Consistent improvements on CODrone and SIMD further demonstrate its robust generalization and efficient performance for aerial vision tasks.