Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (11): 215-223.DOI: 10.3778/j.issn.1002-8331.2309-0177

• Graphics and Image Processing • Previous Articles     Next Articles

Convolutional Image Correction Model for Atmospheric Turbulence Distortion

CHENG Kuanhong, WU Yubo, ZHU Lingjian, LI Jia, LI Junhuai   

  1. 1.School of Computer Science and Engineering, Xi’an University of Technology, Xi’an 710048, China
    2.School of Mechanical and Precision Instrument Engineering, Xi’an University of Technology, Xi’an 710048, China
    3.Department of Basic Sciences, Air Force Engineering University, Xi’an 710051, China
  • Online:2024-06-01 Published:2024-05-31

面向大气湍流畸变场景的卷积图像校正方法

成宽洪,吴钰博,朱凌建,李佳,李军怀   

  1. 1.西安理工大学 计算机科学与工程学院,西安 710048
    2.西安理工大学 机械与精密仪器工程学院,西安 710048
    3.中国人民解放军空军工程大学 基础部,西安 710051

Abstract: The random fluctuations in refractive index caused by atmospheric turbulence in time and space can cause spatiotemporal blur and geometric distortion in images captured by remote imaging systems, seriously weakening the visual effect and application value of the images. In response to this issue, many scholars have attempted to use multi frame lucky zone methods and deep learning methods based on convolutional neural networks to repair image distortion caused by atmospheric turbulence. However, in the case of strong turbulence, these methods usually have greater difficulty in model training and weaker adaptive correction ability for turbulence. To address the above issues, an improved deep learning network has been proposed through research to enhance the performance of turbulence distortion correction. The network adopts a Transformer end-to-end network structure, utilizing multi head self attention to capture local contextual information across channels. At the same time, the first level correction network is trained using a Monte Carlo Dropout strategy to extract degraded regions that are difficult to capture by traditional methods through model uncertainty. Finally, using the extracted uncertainty map as guidance information, input it into the second level correction network to improve the accuracy of correction. Experiments are conducted on a synthetic turbulence degraded image set based on spatiotemporal blur and geometric distortion, demonstrating the effectiveness of the proposed method.

Key words: image processing, atmospheric turbulence, self attention mechanism, uncertainty map

摘要: 大气湍流引发的折射率在时间和空间上的随机波动,会使远程成像系统捕获的图像出现时空模糊和几何畸变,严重削弱了图像的视觉效果和应用价值。针对这一问题,许多学者尝试采用多帧幸运区的方法和基于卷积神经网络的深度学习方法来修复大气湍流引起的图像畸变。然而,在强湍流情况下,这些方法通常模型训练难度较大,湍流自适应性校正能力较为薄弱。为了解决上述问题,经研究提出了一种改进的深度学习网络来提升湍流畸变校正的性能。该网络采用Transformer端到端的网络结构,利用多头自注意力跨通道捕获局部上下文信息;对一级校正网络采用蒙特卡洛Dropout策略进行训练,通过模型不确定性来提取传统方法难以捕获的退化区域;利用提取到的不确定性映射作为引导信息,输入第二级校正网络提升校正准确性。在基于时空模糊加几何畸变的合成湍流退化图像集上进行了实验,证明了提出方法的有效性。

关键词: 图像处理, 大气湍流, 自注意力机制, 不确定性映射