光学精密工程2025，Vol.33Issue(18)：2962-2979,18.DOI:10.37188/OPE.20253318.2962
均值强度差先验驱动的水下图像复原
Underwater image restoration driven by the mean intensity difference prior
摘要
Abstract
Underwater images are degraded by wavelength-dependent absorption and multiple scattering.As a result,they exhibit low contrast,strong color cast,and blurred detail.To address these issues,an underwater image restoration model driven by the mean intensity difference prior(MIDP)was proposed.First,the mean intensity difference between the'blue-green,red'and'green-blue,red'branches was calculated in parallel within each local window and averaged to construct a noise-robust three-channel cou-pled attenuation prior.Subsequently,based on the physical relationship of the'intensity difference-depth map-transmission map'chain,exponential mapping with an adaptive slope was introduced.The three-channel transmission maps were accurately estimated through a first-order Taylor expansion combined with a finite-difference approximation.Meanwhile,the mean of the set of pixels with the largest intensity difference was calculated to estimate the background light.Finally,the transmission maps and the back-ground light were substituted into the inverse underwater imaging model to obtain restored images with nat-ural colors and sharp details.First,image channels were split into two spectral branches:blue-green ver-sus red,and green-blue versus red.Mean intensity differences were computed in parallel for both branches inside every sliding window.The two branch differences were averaged,producing a coupled three-chan-nel attenuation prior for estimating intensity difference.The intensity differences estimated by MIDP were highly robust against image noise and local outliers.Next,a chained physical relation linked intensity dif-ference,scene depth,and transmission map.An exponential mapping with adaptive slope transformed dif-ferences into initial transmission map estimation.A first-order Taylor expansion and finite-difference ap-proximation then refined per-channel transmission maps.Pixels with the largest intensity differences were selected,and their mean was used to estimate background light.The estimated transmission map and background light were inserted into the inverse underwater imaging model to restore clear images.The re-sulting image exhibited natural colors,stronger contrast,and sharper detail.Extensive experiments are conducted on real datasets UIEB and UCCS.Quantitative evaluation shows MIDP obtains the highest combined scores on UIQM,UCIQE,CCF,and MWF.Quantitative evaluation shows that MIDP ob-tains color fidelity,local contrast,and gradient sharpness rise significantly.Additional application experi-ments were carried out on three vision tasks.In key-point detection,the restored images yielded many more stable keypoints.In edge detection,the restored images produced clearer and more coherent re-sults.In underwater segmentation mask,the restored images produce tighter mask boundaries and lower false positive rates.These results demonstrate that MIDP adapts to varied water conditions and restores images with high quality.These results also supply high-quality input data for downstream underwater vi-sion tasks.关键词
水下图像复原/低质图像增强/强度差先验/透射图估计/背景光估计Key words
underwater image restoration/low-quality image enhancement/intensity difference prior/transmission map estimation/background light estimation分类
计算机与自动化引用本文复制引用
商家硕,李颖,袁靖懿,郭深,邢虎..均值强度差先验驱动的水下图像复原[J].光学精密工程,2025,33(18):2962-2979,18.基金项目
国家重点研发计划(No.2024YFB3908800) （No.2024YFB3908800）
中央高校基本科研业务费专项资金(No.3132023507) （No.3132023507）
大连市高层次人才团队创新支持计划(No.2022RG02) （No.2022RG02）
