基于多分支残差注意力网络的水下图像增强

打开文本图片集

中图分类号：TP394. 1；TH691. 9 文献标识码：Adoi：10. 37188/OPE. 20253307. 1141 CSTR：32169. 14. OPE. 20253307. 1141

Underwater image enhancement based on multi-branch residual attention network

CHENG Zhuming*，LI Jiaxuan，HUANG San'ao，HAN Lichao，WANG Peizhen （School of Electrical and Information Engineering，Anhui University of Technology， Maanshan ，China） * Corresponding author， E -mail：czm602@ahut. edu. cn

Abstract：To address color distortion，low contrast，and blurred details in underwater images，a novel en⁃ hancement algorithm based on a multi-branch residual attention network is proposed. Initially，a multibranch color enhancement module is integrated before and after the encoder and decoder to adaptively cor⁃ rect image color deviations. Subsequently，a residual attention module is incorporated at the network’s bottleneck to mitigate feature loss between the encoder and decoder，thereby improving image detail pres⁃ ervation. A composite feature loss function is employed to facilitate comprehensive feature learning and ef⁃ fective retention of edge information. Experimental results demonstrate that the proposed algorithm achieves superior performance in both subjective perception and objective evaluation metrics. Specifically， the average peak signal-to-noise ratio（PSNR）and structural similarity index（SSIM）on the LUSI test set reach 27. 420 dB and 0. 885，representing improvements of 3.9% and 0.8% ，respectively，over the next best method. On the EVUP test set，PSNR and SSIM attain 26. 159 dB and 0. 851，with enhance⁃ ments of 3.3% and 1.3% ，respectively. These results confirm the algorithm's effectiveness and robust⁃ ness in underwater image quality enhancement，offering a valuable approach for image analysis in underwa⁃ ter engineering applications.

Key words：underwater image enhancement；deep learning；residual attention module；multi-branch col⁃ or enhancement module；attention mechanism；joint loss function

1 引 言

在海洋探索、水下生物学研究、水下考古以及水下机器人等领域，通常需要利用水下图像处理技术来帮助捕捉和解析水下世界的视觉信息。（剩余15178字）