【作者】 刘敏；

【导师】 石振锋；

【作者基本信息】 哈尔滨工业大学 ， 数学， 2022， 硕士

【摘要】 由于水下成像环境的复杂,并且受浮沙、浮游生物和光照的影响而动态变化,水下图像往往具有偏色、对比度低、细节信息模糊的特点,原始的水下视频图像很难达到视觉质量的要求。近年来,为了提高水下图像的视觉质量,水下图像增强受到了广泛的关注和深入地研究。随着深度学习在计算机视觉领域的巨大成功,基于深度学习的水下图像增强算法逐渐成为了水下图像领域的研究热点。然而,同时获得大量退化和清晰的配对水下图像是几乎不可能的,缺少训练数据使得深度学习模型的训练变得困难。为此,研究人员采用合成水下图像或人工筛选最佳视觉质量的增强图像作为训练数据。合成水下图像与真实水下图像在逼真度和场景上有着较大的差距,人工筛选增强图像的方式得到的不是真实的参考图像,受主观因素影响可能会导致一些标注上的矛盾。本文针对训练数据不足的问题进行解决,提出了基于域自适应的水下图像增强框架。首先,本文设计了域自适应模型,通过判别学习缩小合成水下图像与真实水下图像在潜在风格上的差异,通过对比学习损失函数保留了合成水下图像的内容信息。在此基础上,本文设计了负样本挖掘模块保证对比学习得到充分利用。然后,本文设计了水下图像增强模型,采用多尺度编码解码模型对水下图像进行增强。最后,本文通过端到端的方式训练域自适应模型和水下图像增强模型。在公开基准数据集上,本文提出的模型性能优于其他的算法。更多还原

【Abstract】 Due to the complexity of the underwater imaging environment and the dynamic changes influenced by floating sand,plankton and light,the underwater images are often characterized by partial color,low contrast and fuzzy details,and the original underwater video images are difficult to meet the requirements of visual quality.In recent years,in order to improve the visual quality of underwater images,underwater image enhancement has received extensive attention and in-depth research.With the great success of deep learning in the field of computer vision,underwater image enhancement algorithm based on deep learning has gradually become a research hotspot in the field of underwater image.However,obtaining a large number of degraded and clear paired underwater images at the same time is nearly impossible,and the lack of training data makes the training of deep learning models difficult.To this end,the researchers used synthetic underwater images or manually screened enhanced images of optimal visual quality as training data.There is a big gap between the synthetic underwater image and the real underwater image in terms of fidelity and scene.The way of manually screening the enhanced image is not the real reference image,which may lead to some labeling contradictions due to subjective factors.In this paper,an underwater image enhancement framework based on domain adaptive is proposed to solve the problem of insufficient training data.Firstly,a domain adaptive model is designed to narrow the potential stylistic differences between synthetic underwater images and real underwater images through discriminant learning,and retain the content information of synthetic underwater images through comparative learning loss function.On this basis,this paper designs a negative sample mining module to ensure that the comparative learning can be fully utilized.Then,the underwater image enhancement model is designed,and multi-scale coding and decoding model is used to enhance the underwater image.Finally,this paper trains the domain adaptive model and underwater image enhancement model in an end-to-end manner.On public benchmark datasets,the model proposed in this paper outperforms other algorithms.更多还原