基于多色域特征与物理模型的水下图像增强

PDF下载HTML

Title:

Underwater image enhancement based on multicolor space features and physical models

作者:

张瑞航, 林森

沈阳理工大学 自动化与电气工程学院, 辽宁 沈阳 110159

Author(s):

ZHANG Ruihang, LIN Sen

School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China

关键词:

水下图像增强; 成像模型; 深度学习; 多色域空间; 特征聚合; 轮换训练; 算法推广; 卷积神经网络

Keywords:

underwater image enhancement; image formation model; deep learning; multicolor space; feature aggregation; alternate training; algorithm generalization; convolutional neural network

分类号:

TP391.41

DOI:

10.11992/tis.202312004

摘要:

水下智能机器人在探测海洋信息时易受悬浮颗粒和光衰减现象的干扰，导致视觉图像退化，造成色彩扭曲、细节模糊等现象。针对上述问题，提出基于多色域特征与物理模型的水下图像增强。首先，设计多色域特征聚合网络，旨在利用不同色域空间提供的信息帮助图像颜色恢复。其次，为获取到更真实的视觉效果，对白平衡算法进行推广，并将深度学习算法与水下光学成像模型结合，以数据驱动的方式求解清晰图像。最后，提出多色域轮换模式对网络进行训练，在不同色域空间中搜索最优解。实验证明，该方法在色彩平衡、细节恢复方面效果显著，相比经典算法与前沿算法更具优势，在特征点匹配与显著性检验任务中满足水下智能机器人视觉系统对图像清晰度的要求。

Abstract:

Underwater intelligent robots are susceptible to interference from suspended particles and light attenuation phenomena when detecting oceanic information, which leads to the degradation of visual images, causing color distortion and blurring of details. An underwater image enhancement method based on multicolor domain features and physical models is proposed to address these issues. First, a multicolor space feature aggregation network is designed to leverage information from different color spaces to aid in color recovery. Then, a generalized white balance algorithm is applied to achieve a more realistic visual performance, and deep learning algorithms are combined with underwater optical imaging models to produce clear images in a data-driven manner. Finally, a multicolor space alternation model is introduced to train the network and optimize parameters across different color spaces. Experiment results demonstrate that this method effectively improves color balance and detail recovery, outperforming classical and novel algorithms. The proposed method meets the image clarity requirements of underwater intelligent robot vision systems in tasks such as feature point matching and saliency detection.