Scholarlink

<p>Small target detection in modern UAV remote sensing images faces challenges such as small target size, complex backgrounds, and blurred features, leading to insufficient detection accuracy in existing algorithms. This paper proposes an improved multi-scale target detection network Sobel-GELAN (Generalized Efficient Layer Aggregation Network) and AIFI-Dilated-Attention-You Only Look Once (SGA-YOLOv8). Firstly, a generalized efficient layer aggregation network incorporating Sobel convolution is proposed to enhance edge feature extraction. Secondly, the Attention-based Intra-scale Feature Interaction (AIFI) module is substituted for the Spatial Pyramid Pooling-Fast (SPPF) module to overcome the feature distortion problem of the SPPF module due to multi-scale pooling. Then, Multi-dimension Augmented Cross-scale Feature Pyramid Network (MAC-FPN) is innovatively designed to synergistically optimize the multi-scale feature fusion and spatial context modeling capabilities. Finally, a feature head Dilated and Attention-based Feature Head (DAFHead) based on dilated convolution and attention mechanism is developed. By increasing the convolution kernel’s sampling interval, dilated convolution extends the receptive field without reducing feature map resolution, thus capturing global context for small targets. Experimental results on the VisDrone-2019 dataset demonstrate that SGA-YOLO outperforms various state-of-the-art detection models, achieving a 13.2% higher mAP@0.5 than the baseline while maintaining competitive computational efficiency, which significantly improves both detection accuracy and robustness.</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

SGA-YOLOv8: Sobel-GELAN and AIFI-dilated-attention fusion network for UAV remote-sensing small target detection

  • 期刊论文

  • 发表日期：

    2025年8月31日

查看全文
  • Shizhou Xu,
  • Kaidi Cui,
  • Guangcong Chen,
  • Mengjie Zhang

摘要

Small target detection in modern UAV remote sensing images faces challenges such as small target size, complex backgrounds, and blurred features, leading to insufficient detection accuracy in existing algorithms. This paper proposes an improved multi-scale target detection network Sobel-GELAN (Generalized Efficient Layer Aggregation Network) and AIFI-Dilated-Attention-You Only Look Once (SGA-YOLOv8). Firstly, a generalized efficient layer aggregation network incorporating Sobel convolution is proposed to enhance edge feature extraction. Secondly, the Attention-based Intra-scale Feature Interaction (AIFI) module is substituted for the Spatial Pyramid Pooling-Fast (SPPF) module to overcome the feature distortion problem of the SPPF module due to multi-scale pooling. Then, Multi-dimension Augmented Cross-scale Feature Pyramid Network (MAC-FPN) is innovatively designed to synergistically optimize the multi-scale feature fusion and spatial context modeling capabilities. Finally, a feature head Dilated and Attention-based Feature Head (DAFHead) based on dilated convolution and attention mechanism is developed. By increasing the convolution kernel’s sampling interval, dilated convolution extends the receptive field without reducing feature map resolution, thus capturing global context for small targets. Experimental results on the VisDrone-2019 dataset demonstrate that SGA-YOLO outperforms various state-of-the-art detection models, achieving a 13.2% higher mAP@0.5 than the baseline while maintaining competitive computational efficiency, which significantly improves both detection accuracy and robustness.