一种基于深度学习的水下扭曲图像复原方法  

A Deep Learning Reconstruction Approach for Underwater Distortion Image

作　　者：敖珺[1] 吴桐 马春波[1] AO Jun;WU Tong;MA Chun-bo(Guilin University of Electronic Technology,School of Information and Communication Engineering,Guilin Guangxi 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《计算机仿真》2020年第8期214-218,共5页Computer Simulation

基　　金：国家自然科学基金(61167006);“认知无线电与信息处理”省部共建教育部重点实验室主任基金(CRKL150106);广西自然科学基金(2018JJA170092);桂林电子科技大学研究生教育创新计划资助项目(2017YJCX25)。

摘　　要：针对水面随机波动对水下目标造成的扭曲畸变问题,提出一种基于深度学习的水下扭曲图像复原算法,通过神经网络学习对应图像间的空间变换关系并进行复原。浮动图像和固定图像经过卷积神经网络后输出局部形变参数,然后经过空间变换网络进行B样条插值得到复原图像。实验结果表明,算法对自然场景下的水下扭曲图像和生成扭曲图像均有明显的校正效果。和传统的迭代复原算法相比,可实现端到端直接输出,运行时间大幅度减少。Aiming at the distortion of underwater objects caused by random fluctuation of water surface,the paper proposed an underwater image restoration algorithm based on deep learning.The spatial transformation relationship between the corresponding images was learned through neural networks and restoration was performed.The floating image and the fixed image were processed through convolutional neural networks to output local deformation parameters.Then B-spline interpolation was performed through the spatial transformer networks to obtain the reconstructed image.Experimental results show that the algorithm has obvious correction effect on underwater distortion images in natural scenes and synthetic distortion images.Compared with the traditional iterative restoration algorithms,the algorithm can output directly from end to end and the running time is greatly reduced.

关 键 词：水下图像 几何畸变 图像配准 卷积神经网络 空间变换网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]