基于多层BP神经网络的无参考视频质量客观评价

No Reference Video Quality Objective Assessment Based on Multilayer BP Neural Network

机器学习在视频质量评价(Video quality assessment,VQA)模型回归方面具有较大的优势,能够较大地提高构建模型的精度.基于此,设计了合理的多层BP神经网络,并以提取的失真视频的内容特征、编解码失真特征、传输失真特征及其视觉感知效应特征参数为输入,通过构建的数据库中的样本对其进行训练学习,构建了一个无参考VQA模型.在模型构建中,首先采用图像的亮度和色度及其视觉感知、图像的灰度梯度期望值、图像的模糊程度、局部对比度、运动矢量及其视觉感知、场景切换特征、比特率、初始时延、单次中断时延、中断频率和中断平均时长共11个特征,来描述影响视频质量的4个主要方面,并对建立的两个视频数据库中的大量视频样本,提取其特征参数;再以该特征参数作为输入,对设计的多层BP神经网络进行训练,从而构建VQA模型;最后,对所提模型进行测试,同时与14种现有的VQA模型进行对比分析,研究其精度、复杂性和泛化性能.实验结果表明:所提模型的精度明显高于其14种现有模型的精度,其最低高出幅度为4.34%;且优于该14种模型的泛化性能,同时复杂性处于该15种模型中的中间水平.综合分析所提模型的精度、泛化性能和复杂性表明,所提模型是一种较好的基于机器学习的VQA模型.

Machine learning has a great advantage in the regression of video quality assessment(VQA)model and can greatly improve the accuracy of built model.To this end,a reasonable BP neural network is designed,and taking the feature values of the distorted video contents,code and decode distortion,transmission distortion,and visual perception effect as inputs,a no reference VQA model is constructed by training them with the samples of the built video databases.In modeling,firstly,11 features are used to describe the four main factors that affect video quality,which are the brightness and chroma of image and their visual perception,the gray gradient expectation of image,the blur degree of image,the local contrast,the motion vectors and their visual perception,the scene switching feature,the bitrate,the initial delay,the single interrupt delay,the interrupt frequency and the average time of interrupt.And the feature parameters of a large number of video samples in the two video databases established are extracted.Then by using these feature parameters as inputs,the BP neural network is trained to construct our VQA model.Finally,the proposed model is tested and compared with 14 existing VQA models to study its accuracy,complexity and generalization performance.The experimental results show that the accuracy of the proposed model is significantly higher than those of 14 existing models,and the lowest increase was 4.34%.And in the generalization performance,it is better than 14 models.Moreover,the complexity of the proposed model is at the intermediate in the 15 VQA methods.Comprehensively analyzing the accuracy,generalization performance and complexity of the proposed model,it is shown that it is a good VQA model based on machine learning.