改进U-Net的全局特征融合水下图像增强网络

打开文本图片集

关键词：水下图像增强；深度学习；特征融合；注意力机制；卷积神经网络中图分类号：TP394.1;TH691.9 文献标识码：Adoi：10.37188/OPE.20263402.0322 CSTR：32169.14.OPE.20263402.0322

Underwater image enhancement network based on improved U-Net with global feature fusion

GAO Shaoshu1，JIAO Guangsen1*，LIGuangfeng1，LIU Zongen²

（1. Qingdao Institute of Software，College of Computer Science and Technology， Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software， China University of Petroleum（East China），Qingdao 266580，China; 2.Petroleum Industry Training Center，China University of Petroleum （East China）， Qingdao 266580，China） * Corresponding author， E-mail： z23070072@s. upc. edu. cn

Abstract：To address thecolor deviation and detail blur of underwater images caused by scattering and attenuation when light propagated in the underwater environment，an improved U-Net global feature fusion underwater image enhancement network was proposed. Firstly，a multi-residual convolution module was designed in theencoder and decoder to fuse the feature information hierarchicallytoreduce the loss of detail information.Secondly，the channel attention module was introduced into the decoder to weight the channels to alleviate the problem of different degrees of channel degradation.Finally，a convolution-per muted self-attention module was designed in the decoder to fuse the global information and promote the network-guided image reconstruction. The proposed method was tested on UIEB dataset，and finally achieved 23.42，0.9OO 5 and O.138 5 on PSNR，SSIM and LPIPS，respectively. The results on LSUI dataset were 29.35，0.938 2 and O.088 O on PSNR， SSIM and LPIPS，respectively. Compared with other commonly used underwater image enhancement methods on several public underwater image datasets，the experimental results show that the proposed method has good effect in restoring color deviation and reducing detail blur，which proves its effectiveness and feasibility.

Key words： underwater image enhancement； deep learning； feature fusion； attention mechanism； convolutional neural network

1引言

清晰的水下图像在各种海洋探索活动中起着至关重要的作用[1-3]，然而，由于光在水下环境中传播时会出现散射和衰减，因此水下图像常受到严重的退化影响，比如颜色偏差和细节模糊[46]。（剩余18808字）