极地研究

极地研究 ›› 2026, Vol. 38 ›› Issue (1): 122-137.DOI: 10.13679/j.jdyj.20240047

所属学科：极地信息与工程技术

• 研究论文 • 上一篇    下一篇

基于注意力和特征融合的渐进式多阶段南极目标体监测图像去噪算法

张宇1,2,4, 窦银科1,2,3, 赵亮亮1, 焦阳阳1, 郭栋梁1   

  1. 1山西省能源互联网研究院, 山西 太原 030032;
    2太原理工大学, 电气与动力工程学院, 山西 太原 030024
    3煤电清洁智能控制教育部重点实验室, 山西 太原 030024;
    4太原工业学院, 自动化系, 山西 太原 030008
  • 收稿日期:2024-04-28 修回日期:2024-07-30 出版日期:2026-03-31 发布日期:2026-04-27
  • 通讯作者: 窦银科
  • 基金资助:
    山西省重点研发计划项目资助

A progressive multi-stage image denoising algorithm for Antarctic target monitoring based on attention and feature fusion

ZHANG Yu1,2,4, DOU Yinke1,2,3, ZHAO Liangliang1, JIAO Yangyang1, GUO Dongliang1   

  1. 1Shanxi Energy Internet Research Institute, Taiyuan 030032, China;
    2College of Electrical and Power Engineering, Taiyuan University of Technology, Taiyuan 030024, China;
    3Key Laboratory of Cleaner Intelligent Control on Coal & Electricity, Ministry of Education, Taiyuan 030024, China;
    4Department of Automation, Taiyuan Institute of Technology, Taiyuan 030008, China
  • Received:2024-04-28 Revised:2024-07-30 Online:2026-03-31 Published:2026-04-27

摘要： 受南极强风雪和强磁场影响, 在南极站区级野外目标体图像监测中, 采集的图像存在自然噪声或内部噪声, 严重影响图像质量, 从而影响监测结果。因此, 本文提出了一种基于注意力和特征融合的渐进式多阶段南极目标体监测图像去噪算法。该算法能够提升目标体图像的清晰度和真实感, 消除剩余的噪声并保留图像的细节和结构, 降低高分辨率特征图的计算复杂度。利用南极现场拍摄的目标体监测图像数据集, 对该算法进行了验证实验。实验结果显示, 监测图像在椒盐噪声、周期噪声以及标准差为70的高斯噪声下, 峰值信噪比和结构相似性分别达到41.82 dB、38.04 dB、37.08 dB和0.991、0.952、0.938, 证明该算法的性能优于主流去噪方法, 同时还具有较低的模型复杂度、更强的抑噪能力和抗干扰性, 为南极现场无人化图像监测技术提供了一种更为可靠的技术手段。

关键词: 南极, 目标体, 图像监测, 深度学习, 多阶段图像去噪

Abstract:

Owing to the influence of intense snow and the magnetic field in Antarctica, the images collected during field object image monitoring at the Antarctic Research Station have natural or internal noise that seriously affects the image quality and thus affects the monitoring results. Therefore, this study proposed a progressive multi-stage image denoising algorithm based on attention and feature fusion to improve the clarity and realism of the image, eliminate the remaining noise and preserve the details and structure of the image, and reduce the computational complexity of high-resolution feature maps. The algorithm was verified using the object body dataset of monitoring images taken in Antarctica. The experimental results showed that the peak signal-to-noise ratio and structure similarity index measure of the monitored images were 41.82 dB, 38.04 dB, 37.08 dB and 0.991, 0.952, 0.938, respectively, in the presence of salt and pepper noise, periodic noise, and Gaussian noise with standard deviation of 70. The algorithm performed better than mainstream denoising methods and had lower model complexity and stronger noise suppression ability and anti-interference ability. Therefore, the algorithm provides a more reliable technical means for managing the unmanned image monitoring technology in the Antarctic research station.


Key words: Antarctic, target object, image monitoring, deep learning, multi-stage image denoising