Network Video Quality Assessment Method Based on Artificial Learning Method

Zhiming Shi ； Chengti Huang

《電腦學刊》 30卷5期 (2019/10) Pp. 172-184

https://doi.org/10.3966/199115992019103005013

摘要

Now, people watch network video in various ways, especially mobile terminals. The service of network video will become the busiest network business. However the quality of network video will be impaired by various factors. So it is very important to monitor the video quality in real time and ensure the service. Many existing objective methods are designed for the specific video distortion, which don't have extensive applicability. Other disadvantage is that the spatial and temporal parameters are not considered simultaneously. Objective methods still need a lot of research to do. In this paper, the artificial learning method is used in the network video quality assessment. It can adjust the objective assessment model, according to the actual network environment. Network and video parameters are comprehensively considered in this method. They are media delivery index (MDI), Noise standard deviation (N_(sd)), Blur degree (B_d), and Block effect (B_e). Every parameter has relationship with the video quality. The contributions of this paper are: (1) The M5' model tree is used to train the network parameters and video parameters. It is an innovation in this domain. (2) Spatial and temporal parameters are considered simultaneously. All the extracted parameters have relationship with the visual perception. (3) The proposed method can improve the accuracy of objective score. In order to validate the proposed method, six videos of different bit-rate are tested under different experimental environment. Firstly 23 people are arranged to watch the network video and give the subjective score. Meanwhile the network and video parameters are extracted from the video. Secondly the M5' model tree is used to model the objective method and train the parameters. Next the objective scores are given and the similarity between the subjective and objective can be computed. Other existing methods are compared with the proposed method. All experimental results show that the proposed method has higher similarity with the subjective assessment. It improves the accuracy of objective score.

關鍵字

artificial learning method ； M5' model tree ； MDI ； similarity ； video parameter ； video quality