基于自然场景统计的无参考视频质量评价算法研究

发布时间：2018-07-29 10:45

【摘要】：视频质量评价在视频应用中有着非常重要的作用,例如,监测视频传输质量、优化视频处理算法和参数等。人眼作为视频的主要接收终端,最准确的视频质量评价方法应该是由人观看并给出评价结果。但是主观评价方法实现流程复杂,耗费时间和人力。因此,人们试图通过客观算法评估视频质量。目前对客观评价算法的研究大多集中在全参考评价方法,尽管性能较好的全参考质量评价算法和主观评价方法结果有较高的一致性,但由于计算中必须获取原始视频的全部信息,使用场景受到限制。无参考视频质量评价算法由于其成本低、实时性高、可移植等特点而受到关注。目前已有一些性能较好的无参考视频质量评价算法,其中由Michele A.Saad等人提出的Video-BLIIND算法性能较好,本文基于该算法对视频质量的评价方法进行研究,完成的主要工作有:1.基于自然时变图像统计分析,提取了和感知质量相关的时空域特征参数作为学习模型中时间域的重要特征。2.对提取出的高维特征进行主成分分析,一方面降低特征参数的维度,另一方面通过主成分分析得出对质量感知影响最大的特征并进行加权处理。3.在Video-BLIIND的基础上,添加视频失真类型判断模块,分两步实现评价算法,第一步判定失真类型,第二步进行分数预测。通过这种设计,将算法转换为已知失真类型的视频质量预测,不仅提高了算法结果的准确性,同时提高了算法的拓展性和灵活性。4.基于视觉掩蔽效应,加入了视频复杂度分析,提取视频的时间复杂度和空间复杂度作为视频质量评价的特征参数。
[Abstract]:Video quality evaluation plays an important role in video applications, such as monitoring video transmission quality, optimizing video processing algorithms and parameters, etc. As the main receiving terminal of video, the most accurate method of video quality evaluation should be viewed by people and given the evaluation results. But the subjective evaluation method is complex, time-consuming and labor-intensive. Therefore, people try to evaluate video quality by objective algorithm. At present, most of the researches on objective evaluation algorithms focus on full reference evaluation methods, although the results of full reference quality evaluation algorithms with better performance are consistent with those of subjective evaluation methods. However, due to the need to obtain all the information of the original video in the computation, the use of the scene is restricted. Non-reference video quality evaluation algorithm has attracted much attention because of its low cost, high real-time and portability. At present, there are some better non-reference video quality evaluation algorithms, among which the Video-BLIIND algorithm proposed by Michele A.Saad et al has better performance. This paper studies the evaluation method of video quality based on this algorithm, and the main work accomplished is: 1. Based on the statistical analysis of natural time-varying images, space-time domain feature parameters related to perceptual quality are extracted as important features of time domain in learning model. The extracted high-dimensional features are analyzed by principal component analysis. On the one hand, the dimension of feature parameters is reduced; on the other hand, through principal component analysis, the features that have the greatest influence on quality perception are obtained and weighted. 3. On the basis of Video-BLIIND, the video distortion type judgment module is added, and the evaluation algorithm is implemented in two steps. The first step determines the distortion type, and the second step carries out the score prediction. Through this design, the algorithm is transformed into video quality prediction with known distortion types, which not only improves the accuracy of the algorithm results, but also improves the expansibility and flexibility of the algorithm. Based on the visual masking effect, video complexity analysis is added to extract the time complexity and space complexity of video as the characteristic parameters of video quality evaluation.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】