水下图像增强技术在促进海洋科学研究、水下考古和资源勘探等领域的发展中发挥着至关重要的作用。然而，由于水下环境的复杂性，获取高质量的水下图像面临诸多挑战，尤其是在现有有标签数据集匮乏且这些数据集标注的图像与对应陆地图像不匹配的情况下，严重限制了图像增强技术的发展和应用。针对这两个问题，本研究提出了两种创新的水下图像增强策略：基于多任务学习的半监督算法和基于双变量扩散模型的增强策略，旨在有效利用无参考图像的水下数据集进行模型训练和图像质量提升。第一种方法，即基于多任务学习的半监督算法，充分利用了无标签水下图像数据的丰富性。本研究采用了均值老师框架作为基础，并结合多任务优化策略，解决了多个有标签水下图像数据集之间分布差异显著的问题。为了确保模型训练的稳定性和评估的准确性，本文还提出了一种新颖的非参考水下图像质量评估方法。该方法基于原始图像与增强图像之间的融合一致性，通过微调预训练的图像质量评估模型，确保了评估结果的准确性和可靠性。此外，为了进一步提高模型的泛化能力和对抗过拟合，我们引入了动量银行队列机制，作为缓冲，使得模型能够更好地从大量的非参考水下真实图像中学习。 第二种方法，即基于双变量扩散模型的增强策略，通过模拟图像质量退化和恢复的动态过程，学习从低质量水下图像到高质量图像的映射。在原本只有一个变量进行逐步加噪的情况下，本文创新性地加入了条件x，通过求解边界条件，确定了双变量扩散过程的形式。进而推导获得了这种基于双变量扩散模型的损失函数，从而能够完成对水下图像增强任务的预训练。通过一系列实验，本研究验证了所提出方法的有效性。在多个评价指标上，文章中提出的的方法均取得了显著的改进，特别是在无参考水下图像的增强任务中。此外，本研究涉及到的多任务学习、半监督框架以及双变量扩散模型的用处非常广泛，为水下图像增强技术的理论发展和实际应用提供了新的视角和强有力的技术支持。未来的工作将集中在进一步优化这些方法，以适应更加多变和复杂的水下环境，推动水下图像增强技术的实际应用和进步。

Underwater image enhancement technology plays a crucial role in advancing various fields such as marine science research, underwater archaeology, and resource exploration. However, obtaining high-quality underwater images faces numerous challenges due to the complexity of the underwater environment, especially in the absence of sufficient labeled datasets as wll as the mismatch between annotated underwater images and corresponding land images, severely limiting the development and application of image enhancement techniques. To address these issues, this study proposes two innovative underwater image enhancement strategies: a semi-supervised algorithm based on multitask learning and an enhancement strategy based on the bivariate diffusion model, aiming to effectively utilize underwater datasets without reference images for model training and image quality improvement.The first method, namely the semi-supervised algorithm based on multitask learning, fully exploits the richness of unlabeled underwater image data. This study adopts the Mean Teacher framework as the foundation and combines multitask optimization strategies to address significant distribution differences among multiple labeled underwater image datasets. To ensure the stability of model training and the accuracy of evaluation, this paper also proposes a novel non-reference underwater image quality assessment method. This method, based on fusion consistency between original and enhanced images, fine-tunes a pre-trained image quality assessment model to ensure the accuracy and reliability of evaluation results. Additionally, to further improve the model‘s generalization ability and combat overfitting, we introduce a momentum bank queue mechanism as a buffer, enabling the model to better learn from a large number of non-reference real underwater images.The second method, namely the enhancement strategy based on the bivariate diffusion model, learns the mapping from low-quality underwater images to high-quality images by simulating the dynamic process of image quality degradation and restoration. In addition to incrementally adding noise to one variable, this paper innovatively introduces a conditional variable x and determines the form of the bivariate diffusion process by solving boundary conditions. Consequently, we derive the loss function of this bivariate diffusion model-based enhancement strategy, thereby completing the pre-training for underwater image enhancement tasks.Through a series of experiments, this study verifies the effectiveness of the proposed methods. Significant improvements are observed in various evaluation metrics, particularly in the enhancement tasks of underwater images without reference. Moreover, the utility of multitask learning, semi-supervised frameworks, and bivariate diffusion models whir are involved in this research is extensive, providing new perspectives and strong technical support for the theoretical development and practical application of underwater image enhancement technology. Future work will focus on further optimizing these methods to adapt to more variable and complex underwater environments, thereby promoting the practical application and advancement of underwater image enhancement technology.