基于视觉信息的质量感知模型及检索方法研究

发布时间：2018-07-18 07:46

【摘要】：随着互联网及计算机视觉的迅猛发展,视觉信息大数据时代也接踵而至。如何从与视觉信息相关的海量的数据中迅速而准确地找到所需信息已成为一个非常有意义且具有挑战性的研究热点,其中又以视觉信息的质量为核心问题。视觉信息的质量直接影响信息的可展示性、检索的准确性与知识发现的可能性,进而影响人工智能推理的可靠性。另外,社交网络上图像、视频等可视化数据的数据量与日俱增,也预示着对多媒体数据有效的检索技术变得越来越重要。因此,本文重点研究视觉信息的统计特性,质量评价准则的设计及基于视觉信息的检索。主要的工作和创新点包括以下几个方面:(1)针对现有基于机器学习的无参考视频质量评价方法中需要利用大量主观评价分值进行训练,导致复杂度高的问题,提出一种非主观值训练的盲视频质量评价算法。该算法的关键是利用高斯差分滤波器提取视频结构特征矢量,并建立多级质量感知中心集合,构建视频空域质量评估密码本;然后利用聚类算法获取对运动矢量进行分类的阈值,进而得到运动感知因子;最后,结合视频空域感知质量和运动加权因子得到视频客观质量。结果表明,该算法优于对比的其他无参视频质量评价算法,且计算复杂度低。(2)综合考虑了影响立体视频用户体验质量QoE的关键因素,提出了一种面向网络传输业务的立体视频用户体验质量的多指标评价模型,该模型首先分析了立体视频QoE的四大类影响因素;然后利用模糊层次分析法对所有影响因素进行层次分析,建立评估指标体系;最后,计算各指标的权重值,得到最终的立体视频QoE评价模型。通过网络仿真实验,并利用主观评测值,验证各个指标对网络传输中立体视频QoE的影响。(3)针对社交媒体中图像和文本之间存在的噪声大及不完全对应,造成基于关键词查找图像的检索方法不准确问题,提出一种基于超图相关性学习的社交图像检索模型。该模型基于超图原理利用图像的视觉信息特征,文本信息,及用户的社交关系信息建立图像之间多类型的关系。在超图的学习过程中,利用选择性优化算法对超边的权重进行更新,实时的调整每个超边在超图重构中的影响。另外,将图像的受欢迎程度用于图像检索的重排序中。该方法适用于在社交媒体互动平台(比如Sina微博),利用图像视觉特征、辅助的文本信息及图像所属用户个人资料和兴趣之间的关系,检索出符合用户个性化要求的结果。
[Abstract]:With the rapid development of Internet and computer vision, the era of visual information big data is also coming. How to find the needed information quickly and accurately from the massive data related to visual information has become a very meaningful and challenging research hotspot, in which the quality of visual information is the core issue. The quality of visual information directly affects the display of information, the accuracy of retrieval and the possibility of knowledge discovery, and then affects the reliability of artificial intelligence reasoning. In addition, the increasing amount of visual data such as images and videos on social networks indicates that effective retrieval techniques for multimedia data are becoming more and more important. Therefore, this paper focuses on the statistical characteristics of visual information, the design of quality evaluation criteria and visual information based retrieval. The main work and innovations include the following aspects: (1) in order to solve the problem of high complexity, we need to use a large number of subjective evaluation scores to train the existing machine learning-based video quality evaluation methods without reference. A blind video quality evaluation algorithm based on non-subjective training is proposed. The key of this algorithm is to extract the feature vector of video structure by using Gao Si differential filter, and to build a set of multi-level quality perception centers and a codebook of video quality evaluation in spatial domain. Then the threshold of motion vector classification is obtained by clustering algorithm, and then the motion perception factor is obtained. Finally, video objective quality is obtained by combining video spatial perception quality and motion weighting factor. The results show that the proposed algorithm is superior to other non-parametric video quality evaluation algorithms, and the computational complexity is low. (2) the key factors affecting the quality of stereoscopic video users' experience quality are considered synthetically. In this paper, a multi-index evaluation model for the user experience quality of stereoscopic video for network transmission service is proposed. Firstly, four kinds of factors affecting the QoS of stereoscopic video are analyzed. Then the fuzzy analytic hierarchy process (FAHP) is used to analyze all the influencing factors, and the evaluation index system is established. Finally, the weight value of each index is calculated, and the final evaluation model of stereoscopic video QoS is obtained. Through the network simulation experiment and the subjective evaluation value, the influence of each index on the QoS of stereo video in the network transmission is verified. (3) aiming at the large noise and incomplete correspondence between images and texts in social media, This paper presents a social image retrieval model based on hypergraph relevance learning. Based on the hypergraph principle, the model uses visual information features, text information and social relationship information of users to establish multi-type relationships between images. In the learning process of hypergraph, the weight of hyper-edge is updated by selective optimization algorithm, and the influence of each hyper-edge in hypergraph reconstruction is adjusted in real time. In addition, the popularity of the image is used in the reordering of image retrieval. This method can be used in social media interactive platform (such as Weibo) to retrieve the results that meet the user's personalized requirements by using the visual features of images, the auxiliary text information and the relationship between the user's personal data and their interests.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】