针对真实雨雾场景的图像质量提升方法研究

Rain streaks and fog have scattering and attenuation effects on light, which decrease image quality and seriously affect visual effects and subsequent processing system performance. Existing methods have the disadvantages of poor generalization ability and low execution efficiency, which make them cannot effectively deal with complex and variable real-world rainy and foggy scenarios, and it is difficult to meet practical application requirements. The project aims to study the combination of non-parametric hierarchical Bayesian theory, deep learning and image processing knowledge to improve the quality of real-world rainy and foggy images. Firstly, a prior estimation scheme based on non-parametric hierarchical Bayesian auto-encoder is proposed to explore the prior information in real-world rainy and foggy images. Secondly, an image quality improvement scheme based on unpaired data training is proposed. The prior information of real-world data is combined with unpaired synthesized data to improve the generalization ability of deep neural networks. This can solve the problem that the existing methods are difficult to effectively deal with real-world rainy and foggy images. Thirdly, expert domain knowledge is introduced to simplify learning problems to design low-memory and high-efficiency neural network structure, and meet the requirements of practical applications for lightweight algorithms. This research not only improves the real-world rainy and foggy image quality, but also provides reference and inspiration for related computer vision problems with practical applications.

雨雾对光线具有散射和衰减作用，导致图像质量降低，严重影响视觉效果与后续处理系统性能。现有方法存在泛化能力较差和执行效率较低等不足，无法有效应对复杂多变的真实雨雾场景，难以满足实际应用需求。项目旨在研究将非参数层级贝叶斯理论、深度学习与图像处理知识相结合的方法，以提升真实雨雾图像的质量。具体为：1）提出基于非参数层级贝叶斯自编码器的真实雨雾图像先验估计方案，挖掘出真实数据中的先验信息；2）提出基于非配对数据训练的图像质量提升方案，将真实数据的先验信息与非配对的仿真数据相结合，提高深度神经网络的泛化能力，解决现有方法难以有效应对真实雨雾图像的问题；3）引入专家领域知识简化学习问题，设计低存储高效率的神经网络结构，满足实际应用对于轻量级算法的要求。相关研究既针对真实雨雾场景的图像质量提升，又可以为面向实际应用的相关计算机视觉问题提供参考和借鉴。

雨天成像环境中的雨线对光具有散射和反射作用，造成图像表观不清、细节模糊等质量退化现象，严重影响后续计算机视觉系统性能。研究真实场景下的图像去雨问题具有重要的理论意义和应用价值。本项目围绕雨天图像复原，重点研究图像复原神经模型建模与学习方法，在神经网络鲁棒表征、轻量化模型构建、模型泛化性学习等方面取得创新进展。项目圆满地完成了预期设定的研究计划，实现了预期目标。对比既定计划，所取得研究成果在深度和广度上都有了进一步的扩展。在项目资助下，发表学术论文34篇，其中IEEE Trans.和CCF A类国际会议长文29篇；申请国家发明专利2项；培养在读研究生2名。