Scholarlink

<p>The identification of the small and blurred targets remains a challenge in the field of computer vision. In this study, we used penguins in images as an example to address this problem. Here we propose an improved YOLOv5-based model (YOLOv5-H) that incorporates a generative adversarial network (GAN) to enhance image quality, introducing a dual-branch convolutional and a CSP for blurred and small objects (CSPBS). We also design a new dynamic convolution module (NDConv) for deep feature extraction. The results show that the parameter count of the YOLOv5-H model is approximately 1.97 times that of YOLOv5s. This substantial increase in model capacity directly improved detection accuracy: precision increased by 14.59%, recall by 74.70%, and mAP@0.5:0.95 by 60.49%. Given the similarities between unmanned aerial vehicles and satellite images, our method could be applied to satellite image processing.</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

Extraction of small, blurred targets using penguins as an example

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • Yankun Hu,
  • Hongyu Chen,
  • Zhaohui Liu,
  • Shaocong Chen,
  • Lijie Ma,
  • Shengjun Gao,
  • Hao Zhang,
  • Ronghao Pan,
  • Xuejiao Wang,
  • Shi Zong,
  • Weixuan Wang,
  • Yingchun Cui

摘要

The identification of the small and blurred targets remains a challenge in the field of computer vision. In this study, we used penguins in images as an example to address this problem. Here we propose an improved YOLOv5-based model (YOLOv5-H) that incorporates a generative adversarial network (GAN) to enhance image quality, introducing a dual-branch convolutional and a CSP for blurred and small objects (CSPBS). We also design a new dynamic convolution module (NDConv) for deep feature extraction. The results show that the parameter count of the YOLOv5-H model is approximately 1.97 times that of YOLOv5s. This substantial increase in model capacity directly improved detection accuracy: precision increased by 14.59%, recall by 74.70%, and mAP@0.5:0.95 by 60.49%. Given the similarities between unmanned aerial vehicles and satellite images, our method could be applied to satellite image processing.