基于改进DeblurGANv2的无人机去模糊目标检测增强方法
作者:
作者单位:

1.国网江苏省电力有限公司无锡供电分公司;2.国网江苏省电力有限公司

中图分类号:

TP273 ????

基金项目:

国网江苏省电力有限公司科技项目(J2023014)


Enhanced Deblurring Target Detection Method for UAVs Based on Improved DeblurGANv2
Author:
Affiliation:

1.Wuxi Power Supply Company of State Grid Jiangsu Electric Power Co.;2.State Grid Jiangsu Electric Power Co., Ltd

  • 摘要
  • | |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • | |
  • 文章评论
    摘要:

    无人机在执行高空巡检作业时,易受风力干扰、机体振动引起图像模糊问题,从而影响目标识别精度。针对于此,提出了一种基于改进DeblurGANv2的无人机去模糊目标检测增强方法:首先,提出了一种基于Haar小波的图像模糊检测方法,通过多级灰度分解与边缘特征提取,实现对模糊区域的高精度检测;其次,设计了一种融合高效通道注意模块与调制变形卷积的改进DeblurGANv2去模糊网络,通过跨通道特征关联与自适应采样策略,实现对模糊特征的有效提取与准确复原。试验结果表明,该方法能有效恢复无人机运动模糊图像的边缘结构与细节特征,相较于DeblurGANv2去模糊方法,复原图像的峰值信噪比提升37.6%,结构相似性提升11.4%。在模糊图像目标检测方面,相对于传统方法识别准确率提升18%,召回率提高30%。

    Abstract:

    Wind disturbances and body vibrations often cause image blurring in unmanned aerial vehicle (UAV) inspections, reducing target recognition accuracy. This study developed an enhanced UAV deblurring and target detection method based on an improved DeblurGANv2. A Haar wavelet-based image blur detection method was introduced, achieving high-precision identification of blurred regions through multi-level grayscale decomposition and edge feature extraction. An improved DeblurGANv2 network integrating an efficient channel attention module and modulated deformable convolution was designed to extract and restore blurred features effectively. Cross-channel feature correlation and an adaptive sampling strategy improved feature extraction and reconstruction accuracy. Experimental results show that the proposed method effectively restores edge structures and fine details in motion-blurred UAV images. Compared to DeblurGANv2, the restored images achieve a 37.6% increase in peak signal-to-noise ratio and an 11.4% improvement in structural similarity. For blurred image target detection, recognition accuracy improves by 18%, and recall increases by 30% over conventional methods.

    参考文献
    相似文献
    引证文献
引用本文
相关视频

分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2025-04-07
  • 最后修改日期:2025-04-14
  • 录用日期:2025-06-11
文章二维码