TECHNICAL
to stay competitive in its markets and seeking assurance it will be able to deliver the QoE it promises to its customers.
more details on this, see Spirent’s accompanying whitepaper, Measuring Video Quality Using AI – Why It’s Relevant and Reliable. As a non-reference type, the algorithm provides these and other insights without requiring access to the source video. This design makes it ideal for all types of streaming video (live and on-demand), all types of providers (vMVPD and SVOD) and all types of networks (wired and wireless). This broad applicability also makes it ideal for video app developers and for manufacturers of endpoints such as smart TVs and players. Although many streaming services started life as low-cost, best-effort alternatives to traditional cable, satellite and telco TV, the market has quickly matured. Now customers expect live and on-demand services to provide a QoE just as good as, if not better than, traditional services. To meet these expectations, the entire streaming ecosystem—from vMVPDs and SVODs to their network operator partners/resellers to endpoint vendors— needs a fundamentally new and better way of measuring QoE. Traditional approaches are based on assumptions that the video must look good because, for example, the packet error rate is below a certain threshold. The combination of AI and non-reference algorithms provides deeper, more actionable insights into the customer experience. This approach to video testing allows organisations to analyse video in the same way a human would perceive it, providing the ability to focus on the actual user experience without the expense of having individuals physically scoring the video content. These insights can also be used for competitors’ networks, enabling providers to benchmark themselves. At the rate the streaming market is maturing and growing, these new tools and insights are must-haves for any provider wanting
University of Southern California. VMAF combines human perceptual vision modeling with AI, and has been shown to be superior to many other algorithms in terms of its ability to produce a score that is well-correlated to how humans rate video quality. An AI VQA system uses several non- reference models, each of which is trained on a specific set of artifacts, such as compression and scaling, which are known to cause problems that customers can see. A variety of VQA models are available to serve as the foundation for a vendor’s set of non-reference algorithms. One example is the Blind/ Referenceless Image Spatial QUality Evaluator (BRISQUE) model, which is a natural-scene, statistic-based blind quality assessment tool developed at the University of Texas at Austin’s Laboratory for Image and Video Engineering (LIVE). BRISQUE is now one of the best, most-used quality assessment tools in broadcast and content production environments, making it ideal for AI VQA systems designed for use with streaming IP video.
Reducing Costs
A growing number of video providers, network operators and other companies are also looking to these AI-based QoE analytics tools for not only quality of experience assessment, but also as a means to help reduce network operating expense. One such example is in use at a major U.S. mobile virtual network operator (MVNO). As a wireless reseller, the MVNO was using video compression to reduce network usage, thus lower the associated fees paid to the mobile network operator. But at the same time, they were concerned with how this compression could adversely impact the mobile user experience. The MVNO now uses the tool to optimise the levels of compression to ensure that they won’t undermine QoE.
Spirent’s Non-Reference Solution
The Spirent Umetrix® Video solution can “view pixels like a person” and score QoE according to VMOS, as if hundreds of human viewers were watching and rating overall performance. Umetrix Video supports any video service (e.g., mobile, home, 5G applications) and analyses the video content via Spirent’s content-trained non-reference algorithm, which uses machine learning on thousands of sample videos to understand the variations in different types of content (sports, drama, animation). Content training is based on de facto industry standards that correlate to human perceptual scoring. The result: faster and less expensive repeatable design validation, regression testing, and competitive benchmarking.
Verifying Accuracy in Identifying Artifacts
Once trained, each non-reference model is tested to see how its scoring compares to an industry standard-scored (in this case, VMAF) video that was not included as part of the initial training. The Pearson correlation is then calculated between the intended score (via VMAF) and non-reference compression model score for each of the thousand compressed clips in the baseline data set. It achieves a correlation of over 90%, which indicates a high level of confidence that it provides an excellent method of scoring video content without using a reference for comparison. For
www.spirent.com
104
SEPTEMBER 2025 Volume 47 No.3
Made with FlippingBook - Online magazine maker