面向盲失真类型的盲视频质量评价关键技术研究

With recent rapid developments of video surveillance and IPTV adopting the large-scale video networks, video quality assessment is a critical research challenge especially when there is no original scene information available. Conventional blind video quality assessment algorithms are mainly designed for tackling the problem of a specific type of distortion incurred in a static image. This project takes more complicated scene factors into consideration, such as interference stripe, illumination, etc. It aims to developing the key technologies of blind video quality assessment for blind distortion types. The information fusion between the spatial-domain and time-domain for dynamic images, as well as the blind identification of distortion model, becomes the critical issue of this study. This project proposes a two-step quality assessment scheme based on example-based training of the natural and distorted video samples. First, through small-scale samples learnt by the support vector machine (SVM) in the feature space, the type of distortion can be identified accurately. The corresponding model parameter can be estimated via the estimated distortion model for blind video quality evaluation. Second, inspired by the fact that the information between the images in a sequence is mutually complementary，a restoration-based method is proposed by using the multiple images to reconstruct a restored image in the feature-domain. This serves as the reference information for the no-reference assessment case to reduce the ill-posed condition of the blind estimation. Third, the ant colony algorithm is used to develop a swarm intelligence approach in the image-domain. After co-evolution of individual intelligence, the swarm intelligence could overcome the limitations of the conventional statistical approaches that need the high accuracy of modeling procedure and estimation complexity. The effectiveness of the proposed approaches will be verified in the video surveillance applications.

随着以视频监控、网络电视为代表的大规模联网，需要仿效人类视觉智能，在缺少原始场景信息的前提下进行无参考的影像质量评价。现阶段盲评价算法主要针对某一特定失真类型的静态图像。本项目考虑更为复杂的场景内容因素，例如干涉、光照等，实现失真类型未知情况下的视频质量盲评价关键技术。动态图像的时空域信息融合和失真模型的盲辨识成为本研究的重点。重点优化以训练学习为指导思想的两步质量评价理论框架。通过支持向量机进行特征空间的小样本学习，对输入视频进行失真类型盲辨识，再选择相应的模型参数估计方法进行盲评价。开展利用多个失真图像间的互补信息进行恢复重建的创新性研究。以变换域空间的重建信息为参考量形成全参考评价方案，降低盲问题的病态程度。以蚁群算法为突破点，通过简单智能个体的协同演化，实现群智能性并摆脱统计方法对退化模型精确性描述及估计束缚。上述理论成果可用于优化相关图像处理算法，并在安防监控领域进行实验验证。

随着以视频监控为代表的大规模联网，监控画面质量受到设备硬件、联网平台、部署环境、人为破坏等不可控因素影响，在视频获取、压缩、传输、存储等环节都会导致视频失真。视频异常原因及类型的多样性，需要仿效人类视觉智能，在缺少原始场景信息的前提下进行无参考的影像质量评价, 使得理论上的视频质量评价方法难以适用。针对重大需求，本项目取得以下原创性成果：.(1)联合相关安防监控企业，项目组收集约30万路实际视频。针对海量视频数据清洗过程中的标签与评测问题，提出采取“3+1+1多级裁判”模式对图像的评判,创建基于Bradley-Terry评价模型的视频质量排序方法，设计和实现了基于偏序方法构建的海量视频数据排序系统，形成视频测试库。.(2)技术面向下一代安防监控，首次提出以训练学习为指导思想的两步质量评价理论框架,重点以深度学习中的卷积神经网络为视频质量评价框架。通过CNN进行图像空间的小样本学习，对输入视频进行失真类型盲辨识，再选择相应的模型参数估计方法进行盲评价。.(3)重点提出和发展基于蚁群演化思想的降质模型盲辨识核心算法。研究发现由于图像噪声相对于周边有剧烈的强度变化，蚂蚁在平滑区域和含噪区域运动路径并不一样。在平滑的区域上，蚂蚁沿着直线向目标运动；反之，运动模式呈现曲线。针对蚂蚁在不同质量视频/图像上的不同运动路径这一特性，完成对失真图像特别是噪声扰动情况下的盲识别。.(4) 新型失真的“内因-表象”模型参数模型，如灰斑异常、树叶遮挡、电线杆遮挡、条纹扰动、偏色干扰、清晰度异常等。理论上首次提出基于流场奇异性分析的图像灰斑检测算法、面向监控视频的树叶遮挡自动判别算法，基于时空域特性的快速条纹扰动等盲评价算法。. 项目执行期间累计发表相关论文30余篇（其中SCI期刊10篇), 2项国家发明专利授权，获批2项软件著作权。上述理论成果不仅用于优化相关图像处理算法，并在安防监控领域进行实验验证，获2017年省科技进步奖二等奖。