Scholarlink

With the growing application of unmanned aerial vehicle (UAV), accurate small object detection in aerial imagery has become increasingly important. However, challenges such as limited object size, occlusion, and poor lighting often result in missed detections and false alarms. To tackle these challenges, we propose DCI-DETR, a small object detection framework with enhanced feature extraction and efficient computation. DCI-DETR incorporates a variant residual network (V-ResNet) that combines Deformable ConvNets v2 (DCNv2) and the parameter-free attention mechanism SimAM to strengthen feature representation. For feature interaction, it replaces the conventional multi-head self-attention (MHSA) in the Attention-based Internal Feature Interaction (AIFI) module with Cascaded Group Attention (CGA), substantially improving efficiency. Furthermore, a novel Shape-Inner Intersection over Union (IoU) loss better aligns bounding box shape and scale, enhancing detection accuracy. Experimental results demonstrate that DCI-DETR significantly enhances performance metrics such as mAP50 and mAP50-95 on the VisDrone2019-DET and CARPK datasets, reflecting excellent generalization and superiority. Code is available at https://github.com/1534736261/DCIDETR .

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

Enhancing Small Object Detection in UAV Aerial Imagery with DCI-DETR: A Deep Learning Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ming Xin,
  • Bo Wang,
  • Xin Wang,
  • Kaiting Gong,
  • Caili Fang

摘要

With the growing application of unmanned aerial vehicle (UAV), accurate small object detection in aerial imagery has become increasingly important. However, challenges such as limited object size, occlusion, and poor lighting often result in missed detections and false alarms. To tackle these challenges, we propose DCI-DETR, a small object detection framework with enhanced feature extraction and efficient computation. DCI-DETR incorporates a variant residual network (V-ResNet) that combines Deformable ConvNets v2 (DCNv2) and the parameter-free attention mechanism SimAM to strengthen feature representation. For feature interaction, it replaces the conventional multi-head self-attention (MHSA) in the Attention-based Internal Feature Interaction (AIFI) module with Cascaded Group Attention (CGA), substantially improving efficiency. Furthermore, a novel Shape-Inner Intersection over Union (IoU) loss better aligns bounding box shape and scale, enhancing detection accuracy. Experimental results demonstrate that DCI-DETR significantly enhances performance metrics such as mAP50 and mAP50-95 on the VisDrone2019-DET and CARPK datasets, reflecting excellent generalization and superiority. Code is available at https://github.com/1534736261/DCIDETR .