Scholarlink

<p>Drone object detection is a critical area within computer vision, facing challenges such as low accuracy in detecting large occluded objects and high rates of missed detection for small objects. These issues arise from inadequate emphasis on the distinctive features of the target area during detection and the regression enhancement of low-quality training examples. To tackle these challenges, this paper introduces a novel method for drone object detection that integrates multi-head mixed self-attention with a dynamic regression mAPping loss function. Initially, a Multi-head Mixed Self-Attention mechanism (MMSA) is developed, tailored to the characteristics of occluded object images. This mechanism is embedded into the backbone and neck components of YOLOv8n to bolster feature extraction and fusion. Subsequently, a dedicated layer for small object detection is incorporated into YOLOv8n to enhance its capability in detecting small objects. A new loss function, Focaler-WIoU, is formulated by merging Focaler-IoU and WIoU, aiming to improve detection across various object scales and accelerate the convergence of bounding box regression loss while enhancing localization accuracy. Additionally, Soft NMS is employed to refine candidate bounding boxes, mitigating missed detections in scenarios with overlapping similar targets. Evaluations on the public dataset VisDrone2019 using standard metrics, including ablation and model comparison experiments, reveal an average precision (mAP&#xa0;@0.5) improvement of 10.2% over the baseline YOLOv8n. The proposed method outperforms other algorithms such as Drone-YOLO (nano), YOLOv11n, Faster RCNN, and FE-YOLOv5 in detection accuracy. Further validation on datasets like CityPersons and CrowdHuman underscores the versatility of the improved algorithm. The experimental outcomes confirm that the MMSA attention mechanism significantly enhances the detection of occluded objects, achieving superior accuracy compared to established object detection algorithms. This suggests that the proposed method holds substantial practical and general applicability for drone image detection in natural settings. Detailed code is available at <a href="https://github.com/CodeSworder/MMSA">https://github.com/CodeSworder/MMSA</a>.</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

Drone object detection incorporating multi-head mixed self-attention and dynamic regression mapping loss function

  • 期刊论文

  • 发表日期：

    2025年2月5日

查看全文
  • Qinghua Su,
  • Jianhong Mu,
  • Sheng Xu,
  • Kaizheng Wan,
  • Xiangyu Qi,
  • Zhichao Zhang,
  • Juntao Li

摘要

Drone object detection is a critical area within computer vision, facing challenges such as low accuracy in detecting large occluded objects and high rates of missed detection for small objects. These issues arise from inadequate emphasis on the distinctive features of the target area during detection and the regression enhancement of low-quality training examples. To tackle these challenges, this paper introduces a novel method for drone object detection that integrates multi-head mixed self-attention with a dynamic regression mAPping loss function. Initially, a Multi-head Mixed Self-Attention mechanism (MMSA) is developed, tailored to the characteristics of occluded object images. This mechanism is embedded into the backbone and neck components of YOLOv8n to bolster feature extraction and fusion. Subsequently, a dedicated layer for small object detection is incorporated into YOLOv8n to enhance its capability in detecting small objects. A new loss function, Focaler-WIoU, is formulated by merging Focaler-IoU and WIoU, aiming to improve detection across various object scales and accelerate the convergence of bounding box regression loss while enhancing localization accuracy. Additionally, Soft NMS is employed to refine candidate bounding boxes, mitigating missed detections in scenarios with overlapping similar targets. Evaluations on the public dataset VisDrone2019 using standard metrics, including ablation and model comparison experiments, reveal an average precision (mAP @0.5) improvement of 10.2% over the baseline YOLOv8n. The proposed method outperforms other algorithms such as Drone-YOLO (nano), YOLOv11n, Faster RCNN, and FE-YOLOv5 in detection accuracy. Further validation on datasets like CityPersons and CrowdHuman underscores the versatility of the improved algorithm. The experimental outcomes confirm that the MMSA attention mechanism significantly enhances the detection of occluded objects, achieving superior accuracy compared to established object detection algorithms. This suggests that the proposed method holds substantial practical and general applicability for drone image detection in natural settings. Detailed code is available at https://github.com/CodeSworder/MMSA.