Video Quality Assessment Model For IPTV Based On Neural Networks

As a result of triple play, Internet Protocol Television (IPTV) has had a rapiddevelopment worldwide. To provide a high-quality service for users, the video qualityshould be assessed by IPTV operators. Objective quality assessment methods have beenwidely used, due to their favorable characteristics such as simple implementation, lowrequirement for manpower and material resources, free from external constrains, and etc.Furthermore, since the parametric bit-stream-layer model of objective qualityassessment method can use the packet-header information and media-related payloadinformation to accurately assess the video quality, it is suitable for IPTV qualityassessment.Aiming at different application scenarios, this thesis investigates two parametricbit-stream-layer models which are suitable for IPTV quality assessment, namely, themodel considering Instantaneous Decoding Refresh (IDR) packet loss rate and themodel based on video content. Since the parameters influencing video quality havecomplex interactions, the two proposed models are both established based on the BackPropagation (BP) neural network which has good characteristics of non-linearprediction and can effectively reflect the mapping between the parameters and the videoquality.The first model is established based on the BP neural network and extracts theinformation of the video quantization parameter (QP), frame rate, packet loss rate andpacket loss rate of IDR from the compressed bit stream to assess the video quality. Thismodel is a parametric bit-stream-layer model with relatively small amount ofcomputation and good real-time property. It can be implemented by IPTV video qualityassessment with low computational capabilities.The video quality is closely related to the video content. The sensitivity of differentvideos to the same packet loss rate is different. In the IPTV video quality assessmentmodel based on video content, the characteristics of video content are considered.Firstly, the information of the QP, bit rate and the motion vector (MV) is extracted fromthe video bit stream. Then the temporal complexity is calculated using the extracted MV,and the spatial complexity using the QP and bit rate. And then we divide the video intoseveral types by cluster analysis method. Finally, the model is established based on theBP neural network to assess the video quality. This method is simple and effective. Andthe model is fit for those IPTV video quality assessments with certain computational capabilities.