采用多尺度密集残差网络的水下图像增强  

Underwater Image Enhancement Using Multi-scale Dense Residual Network

作　　者：卫依雪 周冬明[1] 王长城 李淼 WEI Yixue;ZHOU Dongming;WANG Changcheng;LI Miao(School of Information Science and Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,云南昆明650504

出　　处：《无线电工程》2021年第9期870-878,共9页Radio Engineering

基　　金：国家自然科学基金资助项目(62066047,61365001,61463052)。

摘　　要：为了有效解决水下图像亮度、对比度过低和颜色混乱等问题,提出一种改进的多尺度密集残差网络的水下图像增强方法。对原始图像进行多尺度特征提取,更好地保留了图像细节,通过改进的密集残差网络对水下图像进行增强处理,提升图像亮度和对比度,校正图像颜色,在每个密集残差网络间添加了SK注意力机制,可以选择性地捕捉输入图像的关键信息并进行处理,将增强后的水下图像进行多尺度融合。通过Type和EUVP两个水下图像数据集对所提出方法进行验证,基于物理模型和数据驱动的6种方法进行了主观效果和客观指标间的比较。在主观效果的定性分析中发现,所提出的方法在提高亮度和对比度方面取得了很大的进步。在客观图像评价指标的定量分析中,峰值信噪比(Peak Signal to Noise Ratio,PSNR)、结构相似性(Structural Similarity,SSIM)、信噪比(Signal to Noise Ratio,SNR)、均方误差(Mean Square Error,MSE)、视觉信息保真度(Visual Information Fidelity,VIF)、信息保真度准则(Information Fidelity Criterion,IFC)、噪声质量评价(Noise Quality Measure,NQM)、亮度顺序误差(Lightness Order Error,LOE)和自然图像质量评价(Natural Image Quality Evaluator,NIQE)指标较现有的水下图像增强算法分别提高了1.5%,1%,1.2%,1.2%,1.3%,1.2%,1.7%,3.0%和1.1%。提出的改进多尺度密集残差网络不仅可以增强图像的亮度、对比度以及校正图像的颜色,而且可以应用于更广泛的水域场景。In order to effectively solve the problems of low brightness,low contrast and color confusion of underwater images,an improved multi-scale dense residual network for underwater image enhancement is proposed.Firstly,the original image is subjected to multi-scale feature extraction to better preserve the image details,and then the underwater image is enhanced by improving dense intensive residual network to enhance the image brightness and contrast,and correct the image color.In addition,SK attention mechanism,added between every two dense residual networks,can selectively capture and process the key information of the input image.Finally,multi-scale fusion of the enhanced underwater image is conducted.The proposed method is tested and demonstrated by two underwater image datasets(Type and EUVP),and six methods based on physical models and data-driven are compared between subjective effects and objective metrics.In the qualitative analysis of the subjective effect,it is found that the proposed method has made great progress in the enhancement of brightness and contrast.In the quantitative analysis of objective image evaluation indexes,it is found that peak signal to noise ratio(PSNR),structural similarity(SSIM),signal to noise ratio(SNR),mean square error(MSE),visual information fidelity(VIF),information fidelity criterion(IFC),noise quality measure(NQM),lightness order error(LOE)and natural image quality evaluator(NIQE)indexes are averagely improved by 1.5%,1%,1.2%,1.2%,1.3%,1.2%,1.7%,3.0%and 1.1%respectively,compared with the comparison of the existing underwater image enhancement algorithms.The proposed improved multi-scale dense residual network can not only enhance the brightness and contrast of the images,but also correct the color of the image,and can be applied to a wider range of water scenes.

关 键 词：水下图像 密集残差网络 多尺度 神经网络 图像增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]