基于生成对抗网络的水下图像增强  

Underwater image enhancement based on generative adversarial networks

作　　者：丁元明[1] 刘苏睿 杨阳[1,2] DING Yuan-ming;LIU Su-rui;YANG Yang(Key Laboratory of Communication and Network,Dalian University,Dalian 116622,China;College of Information Engineering,Dalian University,Dalian 116622,China)

机构地区：[1]大连大学通信与网络重点实验室,辽宁大连116622 [2]大连大学信息工程学院,辽宁大连116622

出　　处：《舰船科学技术》2023年第22期143-147,共5页Ship Science and Technology

基　　金：国家自然科学基金资助项目(61901079)。

摘　　要：对光的散射和衰减导致水下图像出现颜色失真和细节模糊对比度低问题进行研究,提出一种基于生成对抗网络(GAN)的图像增强方法。首先以图像分割(U-Net)网络为基础提取水下退化图像特征,再使用改进的白平衡算法对原始图像进行去偏色处理,用卷积神经网络提取去偏色后的图像特征,接着通过卷积神经网络完成两者特征融合,最后重构增强的图像。结果表明,本文算法增强后的图像在UIQM、PSNR和SSIM指标上的平均值为5.071、25.310和0.996,分别比第二名提升了1%、7%和5%。在主观感知和客观评估中,处理后的图像在清晰度、颜色校正和对比度方面均得到改善。Due to light scattering and attenuation,underwater images will suffer from color distortion,blurred details and low contrast.An image enhancement method based on generative adversarial network is proposed.First,extract the features of the underwater degraded image based on the U-Net network,use the improved white balance algorithm to decolorize the original image,use the convolutional neural network to extract the decolorized image features,and then use the convolutional neural network.The network completes feature fusion and finally reconstructs the enhanced image.The average values of this paper on the UIQM,PSNR and SSIM indicators were 5.071,25.310 and 0.996,which were 1%,7%and 5%higher than the second place,respectively.In both subjective perception and objective assessment,the processed image is improved in terms of sharpness,color correction,and contrast.

关 键 词：水下图像增强 白平衡 生成对抗网络 U-Net 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]