面向室外视频监控的图像增强技术研究

Video Surveillance has become an important means for ensuring the national security and maintaining the social stability. However, video surveillance systems are easily affected by extreme weathers, such as haze, rain, and nighttime, which will result in image quality degradation, and objective recognition inaccuracy. The extreme weather seriously limits the usages of video surveillances in practical applications. Therefore, according to different quality degradation reasons and image enhancement requirement of different processes of video surveillance, this project proposes a series image enhancement aiming at different kinds of extreme weather. A single image dehazing algorithm based on accurate scattered light estimation will be proposed to remove haze, and meanwhile, ensure a clear texture of image. A naturalness preserved image enhancement for rainy weather will be proposed to not only enhance the details of image but also preserve the naturalness of images. By considering both the characteristic of video acquisition and the features of human vision system, a video de-noising algorithm for nighttime will be proposed by jointing spatial and temporal correlations of videos. This de-noising algorithm can not only efficiently remove the noise but also maintain temporal smoothness among neighboring frames. Moreover, a feature-level image enhancement aiming at video content analysis is proposed for the first time, which can break through the limitation of conventional image enhancements aiming at improve the human perceptual quality. The proposed feature-level image enhancement can suppress complex background interference and enhance the saliency of the important features, which will be helpful for the following image content analysis. The result of this project can effectively improve the radius of monitoring and the accuracy of objective recognition, which can improve the capability of video surveillance for resisting disturbance from the extreme weathers.

视频监控已成为保障国家安全、维护社会稳定的重要手段。然而，视频监控系统易受雾霾、阴雨和夜间等恶劣天候影响，导致获取图像“看不清”、目标识别“判不准”等问题，制约了在实际应用中的使用效能。为此，本项目针对光学成像、设备采集和计算机分析过程中，不同的图像降质特点及处理需求，分别开展图像增强技术研究。针对雾霾天和阴雨天图像的降质原因，提出基于散射光准确估计的图像去雾算法和自然度保持的非均匀图像增强算法，在提高图像清晰度和增强纹理的同时，保持图像的自然度。综合考虑视频采集特点和人眼视觉特性，提出联合时空相关性的夜间视频去噪算法，在去噪的同时保证视频播放的连续性。在此基础上，突破传统图像增强方法仅能满足“人眼观看”需求的局限，提出面向视频内容分析的特征级增强算法，以抑制复杂环境干扰并提高关键特征的显著性。研究成果将有效提高视频监控系统对恶劣天候的自适应能力，进而解决“看不清”、“判不准”等难题。

视频监控已成为保障国家安全、维护社会稳定的重要手段。然而，视频监控系统易受雾霾、阴雨和夜间等恶劣天候影响，导致获取图像“看不清”、目标识别“判不准”等问题，制约了在实际应用中的使用效能。.因此，本项目旨在研究面向室外视频监控的图像增强技术。本项目构建了完善的成像降质模型，准确刻画了非均匀光照环境下的混合成像降质过程，为浓雾和阴雨天等恶劣环境下清晰成像奠定了理论基础。提出了一种消除粒子散射的雾天图像去雾算法和自然度保持的光照不均匀图像增强算法，实现了天候自适应的多路高清视频清晰化处理。此外，提出了一种多模融合的低照度视频清晰成像方法，实现了低照度下的视频彩色清晰成像，进而解决了恶劣天候下“看不清”的问题。本项目提出了一种面向视频内容分析的特征增强方法，通过域迁移方法来消除降质图像在特征空间上的分布差异，增强目标表观特征和深度特征，进而提升了运动目标检测等视频内容分析任务的准确性，为解决成像质量下降而导致的视频内容分析“判不准”问题提供了支撑。基于上述工作，本项目研发了天时天候自适应视频清晰化处理系统，已在智慧交通和安防监控等领域部署和应用。.本项目累计在国际期刊和国际会议发表学术论文14篇（其中IEEE 汇刊论文9篇），申请国家发明专利7项（授权4项），指导培养研究生7名。此外，项目负责人获批了国家自然基金委优秀青年科学基金项目，并获得了2019年度国家科技进步二等奖（项目负责人排名3/10）。