【作者】 许凯华；

【导师】 李鑫滨； 田新华； Wang Xiaoming；

【作者基本信息】 燕山大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 在海洋资源的开发过程中,水下图像作为人类获取海洋信息的主要手段受到越来越多的重视。水下图像采集与处理技术在海洋探测、水下考古、海洋生物研究等领域扮演着愈发重要的角色。然而,由于水下环境复杂多变,光线在水中传播时受到吸收和散射的干扰,导致采集到的水下图像往往存在颜色失真、对比度降低、细节模糊等问题。这些问题极大地影响了图像信息的准确提取和后续分析。因此,研究水下图像增强技术,提升水下图像质量,具有实用价值和研究意义。本文针对水下图像增强技术研究,提出了适应不同环境需求的水下图像增强方法和多水域环境数据集,具体工作内容如下:(1)本文介绍了传统水下图像增强方法和基于深度学习的水下图像增强方法,阐明了目前水下图像增强任务的难点,对水下图像成像模型、卷积神经网络和生成对抗网络进行深入分析。进行了卷积神经网络和生成对抗网络的水下图像增强对比实验,选取了两个代表性的方法,分别是UWCNN和GAN-RS。通过对实验结果的对比和分析,生成对抗网络在处理图像时展现了更优越的性能且采用无监督的方式具有灵活性,所以选取生成对抗网络作为基本网络框架。(2)为了解决传统水下图像增强方法处理效果不足的问题,提出了一种基于自注意力机制的水下图像增强算法。同时,考虑到水下图像细节失真问题,在生成器结构中添加了自注意力层,增强提取特征能力并润色细节。进而,针对特征图在经过各个层的计算后会丢失细节信息问题,设计了细节增强模块加入到生成器中,弥补出现的特征损失。最后,针对使用合成图像和真实图像之间固有的领域差距问题,在网络中引入了域自适应机制,使生成图像的真实感更接近空中图像。通过消融实验和与其他算法的比较,所提算法有效的提高了水下图像的质量。(3)为了解决采用单一水下环境数据集的增强方法不能适用于水下多变环境的问题,提出了一种基于双重约束的水下图像增强方法。针对现有数据集只包含单一水下环境问题,创建了一个模拟多种水域环境的合成水下数据集,增强网络经过该数据集训练能够适应复杂多样的水下环境。考虑到使用生成对抗网络的方法不能得到物理上真实的图像问题,在框架中添加物理模型作为物理约束,结合框架中的循环约束组成双重约束,并设计了双重约束损失函数。最后,针对生成图像不能够保持三维信息问题,提出了传输损失函数,促进了生成图像在三维深度空间与原始图像相似。实验结果表明,与其它图像增强方法相比,所提方法在主观和客观评价上的指标结果最优,能够更好地满足不同水下环境下的需求。更多还原

【Abstract】 In the process of developing marine resources,underwater images have received more and more attention as the main means for humans to obtain marine information.Underwater image acquisition and processing technology plays an increasingly important role in the fields of ocean exploration,underwater archeology,and marine biology research.However,due to the complex and changeable underwater environment,light is interfered by absorption and scattering when propagating in water,resulting in problems such as color distortion,reduced contrast,and blurred details in collected underwater images.These problems greatly affect the accurate extraction and subsequent analysis of image information.The investigation of underwater image enhancement technology and the augmentation of underwater image quality possess practical utility and research relevance.The present study is devoted to the examination of underwater image enhancement technology,introducing methods for the enhancement of underwater images and datasets for various water environments,tailored to a range of environmental necessities.The specific work content is as follows:(1)The present study presents traditional methods of underwater image enhancement and deep learning-based underwater image enhancement techniques,elucidates the challenges inherent in current underwater image enhancement tasks,and undertakes research on underwater image imaging models,convolutional neural networks,and generative adversarial networks.In-depth analysis.We conducted underwater image enhancement comparison experiments with convolutional neural networks and generative adversarial networks,and selected two representative methods,namely UWCNN and GAN-RS.Through comparison and analysis of experimental results,the generative adversarial network shows superior performance when processing images and is flexible in an unsupervised manner,so the generative adversarial network is selected as the basic network framework.(2)To address the issue of inadequate processing efficacy of traditional underwater image enhancement techniques,the study proposes an underwater image enhancement algorithm grounded in the self-attention mechanism.Considering the problem of underwater image detail distortion,a self-attention layer is added to the generator structure to enhance the ability to extract features and polish details.Furthermore,in order to solve the problem that the feature map will lose detailed information after the calculation of each layer,a detail enhancement module is designed and added to the generator structure to make up for the loss of features.Finally,to address the inherent domain gap problem between using synthetic images and real images,a domain adaptation mechanism is introduced in the network to make the realism of the generated images closer to aerial images.Through ablation experiments and juxtapositions with alternative algorithms,the proposed algorithm demonstrates effective enhancement of underwater image quality.(3)To address the issue that enhancement techniques utilizing a singular underwater environment dataset are not adaptable to varying underwater environments,the study introduces an underwater image enhancement method predicated on dual constraints.Aiming at the problem that the existing dataset only contains a single underwater environment,a synthetic underwater dataset is created that simulates a variety of water environments.The enhanced network can adapt to complex and diverse underwater environments after training with this dataset.Considering that the method of using generative adversarial networks cannot obtain physically realistic images,a physical model is added to the framework as a physical constraint,combined with the loop constraints in the framework to form a double constraint,and a double constraint loss function is designed.Finally,to solve the problem that the generated image cannot maintain three-dimensional information,a transmission loss function is proposed,which promotes the generated image to be similar to the original image in the three-dimensional depth space.Experimental results show that compared with other image enhancement methods,the proposed method has the best index results in qualitative and quantitative analysis,and can better meet the needs of different underwater environments.更多还原