基于密集残差策略的多尺度水下图像增强网络  

Multi-Scale Underwater Image Enhancement Network Based on Dense Residual Strategy

作　　者：李祯 周冬明[1] 周联敏 尹稳 赵倩 LI Zhen;ZHOU Dongming;ZHOU Lianmin;YIN Wen;ZHAO Qian(School of Information Science and Engineering,Yunnan University,Kunming Yunnan 650091,China)

机构地区：[1]云南大学信息学院,云南昆明650091

出　　处：《电子器件》2025年第1期105-115,共11页Chinese Journal of Electron Devices

基　　金：国家自然科学基金项目(62066047,61966037)。

摘　　要：在水下成像过程中,图像往往会产生颜色失真、对比度低等问题。针对以上问题,提出一种基于密集残差策略的多尺度水下图像增强网络。通过构建基于非线性无激活函数模块的U型主干结构,改善了由于网络复杂程度提升所导致的模型训练以及预测耗费大量时间的问题;为了补偿上下采样导致的细节丢失,采用选择性密集融合模块,以大量的残差连接代替卷积堆叠,充分扩宽网络深度,有效提取其中的关键特征;最后,为充分考虑图像在不同阶段的详细特征,引入多尺度融合模块,聚合不同尺度的特征图。结果表明,提出的模型比目前主流的水下图像增强方法在峰值信噪比、结构相似性、水下图像质量评价指标上高,分别达到31.8691 dB、0.8568、2.87,且更符合人眼对图像视觉的要求。In the process of underwater imaging,the image often produces color distortion,low contrast and other problems.To solve the above problems,a multi-scale underwater image enhancement network based on dense residual strategy is proposed.By constructing a U-shaped backbone structure based on nonlinear non-activation function module,the problem of model training and prediction taking a lot of time due to the increase of network complexity is improved.In order to compensate for the loss of details caused by upper and low-er sampling,a selected intensive fusion module is adopted to replace the convolution stack with a large number of residual connections,fully expand the network depth,and effectively extract the key features into the backbone structure.Finally,in order to fully consider the detailed features of the image at different stages,a multi-scale fusion module is introduced to aggregate the feature maps of different scales.The results show that the proposed model is higher than the current mainstream underwater image enhancement methods in terms of peak signal-to-noise ratio,structural similarity and underwater image quality evaluation indicators,reaching 31.8691 dB,0.8568 and 2.87 respectively,and more in line with the requirements of human eyes for image vision.

关 键 词：水下图像 多尺度融合 密集残差网络 图像增强 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TN911.73[自动化与计算机技术—控制科学与工程]