雾霾天气条件下监控视频的复原与增强研究

The foggy and hazy weather, illumination changes and noise are the main obstacles in the technological developments and applications of outdoor surveillance video. This project focuses on improving the theories and methods of the restoration and enhancement of surveillance video under hazy weather conditions. It is devoted to the studies of the noise model of imaging devices, the physical model of hazy video, and point spread model of optical imaging to establish a general degradation model. This general model can accurately describe the quality of surveillance video degraded by hazy weather conditions. According to key points of each processing link and signal characteristics, this project will explore the structural self-similarities, the sparsities and other prior knowledge of surveillance video. Several advanced techniques will be used to extract the component of interest from surveillance video, which include data classification, transform filtering and so on. The interesting component will be attenuated or enhanced to improve the surveillance video quality based on both the optical imaging principle and the human visual perception features. This project focuses on the studies of video denoising based on the noise model of imaging devices, video dehazing based on the optimized physical model, and video enhancement based on both illumination invariance and human visual perception characteristics. This task is also devoted to how to use the numerical simulation and experiments to validate the proposed methods. The validation indicates that the proposed method can successfully increase the surveillance video quality of outdoor high-resolution imaging in all weather conditions. This project tackles the key problems in surveillance video processing technology, and the results can greatly promote the applications of surveillance video in the intelligent transportation and security systems.

雾霾天气、光照变化和噪声是户外监控视频技术发展和应用的主要障碍。本项目将致力于修正和完善雾霾天气下监控视频复原与增强的理论和方法。深入研究成像设备的噪声模型、雾霾视频退化的物理模型和光学成像的点扩散模型，建立能够准确描述雾霾天气下降质监控视频的统一退化模型。充分利用监控视频的结构自相似性和稀疏性等先验知识，按照各处理环节的重点和信号特点，通过数据分类、变换滤波等先进技术，提取视频中感兴趣的成分，根据光学成像原理和人眼视觉感知特性，抑制噪声和雾霾，增强色彩和纹理，改善监控视频质量。具体研究：基于成像设备噪声模型的视频去噪方法，基于修正物理成像模型的视频去雾霾方法，基于光照不变性和视觉感知特性的视频增强方法。利用仿真和实验验证所提户外监控视频复原与增强方法，实现全天候高分辨率外场成像。本项目研究结果将极大改善监控视频质量，并促进监控视频在智能交通和安全防范等领域中的应用。

本项目致力于修正和完善雾霾天气下监控视频复原与增强的理论和方法。深入研究成像设备的噪声模型、雾霾视频退化的物理模型和光学成像的点扩散模型，建立能够准确描述雾霾天气下降质监控视频的统一退化模型。根据光学成像原理和人眼视觉感知特性，研究视频图像的噪声去除，图像去雾霾和图像的超分辨率增强重建等问题，改善监控视频质量。具体进行了如下研究： .针对视频图像采集中的噪声干扰，提出基于主成分分析的分块视频噪声估计算法；将双域滤波和三维块匹配算法相结合，进行视频去噪；引入了形状感知核函数的核Wiener滤波模型来描述局部复杂图像结构进行去噪，能有效地去除噪声。针对现有图像去雾算法存在的天空色彩失真等问题，基于天空分割的暗通道理论的改进去雾算法；并设计了一种能够直接学习和估计有雾图像和霾层图像之间的映射关系的网络模型，使得恢复出的无雾图像更接近真实场景。提出了一种基于景深和稀疏编码的图像去雨算法，使得图像去雨效果更好。针对基于插值的图像超分辨引入噪声的问题，根据图像梯度自适应调整冲击滤波的权重；并利用奇异值分解及阈值化去噪提高局部自相似性提取高频信息的准确性；通过局部自相似性在原始低分辨图像中估计高频信息；并研究了基于卷积神经网络的图像超分辨率算法效果。利用各向异性热扩散方程来估计图像深度信息，研究了基于深度信息提取的多聚焦图像的融合算法；并研究了块结构非相干字典学习算法，用于图像信号的稀疏表示；对于双目立体匹配中的视差计算问题，优化了视差斜平面拟合过程；提出一种基于“缩小型”网络的CNN立体匹配方法，算法在精度和速度方面有一定优势。.本项目按原定计划开展了研究，并取得了预期的研究成果，已完成了预期的考核指标。 迄今为止，本项目在国内外核心期刊发表研究论文全文 33 篇， SCI 检索收录 10 篇； 获批国家发明专利 6 项； 邀请海外专家短期讲学交流 1 次，参加国际和国内学术会议专题报告 12 人次。