南京林业大学学报(自然科学版) ›› 2024, Vol. 48 ›› Issue (2): 175-181.doi: 10.12302/j.issn.1000-2006.202203011

• 研究论文 • 上一篇    下一篇

基于Resnet的林地无人机图像去雾改进算法

牛弘健1,2(), 刘文萍1,2,*(), 陈日强1,2, 宗世祥3, 骆有庆3   

  1. 1.北京林业大学信息学院,北京 100083
    2.国家林业和草原局林业智能信息处理工程技术研究中心,北京 100083
    3.北京林业大学林学院,北京 100083
  • 收稿日期:2022-03-03 修回日期:2022-05-14 出版日期:2024-03-30 发布日期:2024-04-08
  • 通讯作者: *刘文萍(wendyl@vip.163.com),教授。
  • 作者简介:

    牛弘健(niuhj1996@gmail.com)。

  • 基金资助:
    国家重点研发计划(2021YFD1400900);国家林业和草原局重大应急科技项目(ZD202001)

Dehaze algorithm for woodland UAV images based on Resnet

NIU Hongjian1,2(), LIU Wenping1,2,*(), CHEN Riqiang1,2, ZONG Shixiang3, LUO Youqing3   

  1. 1. College of Information Science and Technology, Beijing Forestry University, Beijing 100083, China
    2. National Forestry and Grassland Administration, Forestry Intelligent Information Processing Engineering and Technology Research Center, Beijing 100083, China
    3. College of Forestry, Beijing Forestry University, Beijing 100083, China
  • Received:2022-03-03 Revised:2022-05-14 Online:2024-03-30 Published:2024-04-08

摘要:

【目的】 针对雾霾天气下林地无人机航拍图像存在对比度低、饱和度低和色调偏移等现象,基于Resnet网络,提出一种适应林地航拍场景的无人机图像去雾方法(DHnet)。【方法】 林地场景下无人机图像具有纹理特征、高低频信息丰富的特点,在主干网络各个层级附加信息传递模块,将特征图转化为权值图进行筛选过滤并发送到其他层级,接收端设置阈值避免冗余信息的不良影响,再经密集链接增强全局去雾效果,提高图像高低频区域的去雾质量,最后在林地无人机有雾图像测试集上进行去雾实验。【结果】 DHnet在林地图像测试集上的平均结构相似性为0.83,平均峰值信噪比为22.3 dB,分别较Resnet方法提高了4.8%和39.3%。【结论】 本研究提出的算法能有效降低图像色调偏移,去除残留雾气信息,有效提高无人机航拍林地雾气图像的色彩保真度和细节信息保持度。

关键词: 林地, 无人机, 图像去雾, 深度学习

Abstract:

【Objective】 Aiming to address the phenomena of low contrast, low saturation, and hue shift in unmanned aerial vehicle (UAV) photography images of forestland under hazy conditions, this study proposes a de-fogging method for UAV images adapted to forestland aerial photography scenes based on Resnet. 【Method】 The UAV images in woodland scenes were characterized by texture features and rich high-and low-frequency information. GFF information transfer modules were attached to each layer of the backbone network to transform feature maps into weight maps for filtering and sending to other layers, and thresholds were set at the receiving end to avoid the adverse effects of redundant information. Then, the global defogging effect was enhanced by dense links to improve the defogging quality in high- and low-frequency image regions. Finally, defogging experiments were conducted on a test set of woodland UAV images with fog. 【Result】 The average structural similarity of DHnet on the test set of woodland images was 0.83, and the average peak signal-to-noise ratio was 22.3 dB, which represented improvements of 4.8% and 39.3%, respectively, compared with the Resnet method. 【Conclusion】 The algorithm can effectively reduce tonal shift and remove residual fog, improving the color fidelity and detailed information retention of aerial woodland fog images obtained by UAV photography.

Key words: woodland, unmanned aerial vehicle(UAV), image dehaze, deep learning

中图分类号: