TECHNICAL
Compared to human scoring, such as those obtained in focus groups, AI VQA is objective, faster, far more repeatable and less expensive. AI VQA also enables the use of a non-reference algorithm and a subset of AI, called machine learning, which makes it possible to train the algorithm on thousands of diverse video samples. This process enables it to provide a VMOS based on de facto industry standards and correlated to human perceptual scoring. The scoring is accurate with all of the most widely used resolutions and frame rates, from 480p through 4K, and from 24 to 60FPS. How Non-Reference Algorithms are Trained In an AI VQA that supports non- reference algorithms, the machine learning process feeds the algorithm hundreds of thousands of sample video clips—far more than any human could sit through. Each clip is accompanied by a video quality score based on a respected industry standard. This enables the algorithm to build up a working knowledge akin to “If I see this, then the score must be that.” Once it’s been fully trained, the algorithm is now capable of scoring video that it’s never seen before, whether that’s an on-demand movie or a live newscast. Hence the term “non- reference.” One of the best video scoring systems in the industry (and what Spirent uses) is Video Multimethod Assessment Fusion (VMAF), developed by Netflix and the
of third-party streaming services on its network and then compare that to the QoE on rival providers. This capability is particularly valuable for providers that are using cellular or other wireless technologies, which are subject to more anomalies than copper or fibre. As a result, providers can benchmark their services against those of their wired and wireless competitors to understand how their own video services’ QoE compares to the rest of the marketplace. Video app developers also benefit because they can use lab tests to see if changes to device hardware, firmware and software affect QoE.
Use an Algorithm to Compare Pixels
This approach is similar to the previous one because it compares the received video to the source. This comparison uses full reference algorithms such as Peak Signal-to-Noise Ratio (PSNR), which roughly takes the RGB value difference for every pixel and uses that to compute a score. Another, more sophisticated algorithm is Perceptual Evaluation of Video Quality (PEVQ) which uses models that replicate human vision to evaluate blockiness, blur, noise and other attributes. PEVQ is the better choice for creating a video mean opinion score (VMOS) that correlates with how real people score videos. But the drawback of PEVQ and other similar algorithms is that the test requires access to the source video.
AI Enables the Epitome of Video QoE Analysis
The artificial intelligence (AI) field has advanced to the point that it’s now capable of providing the kind of automated video quality analysis (VQA) that content providers, network operators, endpoint vendors, systems designers and others have sought for decades. It takes an holistic, end-to-end view and enables a wide variety of testing scenarios, such as testing a prototype mobile phone or streaming player to analyse the video QoE it provides on multiple types of network technologies from multiple service providers. Another example is using AI VQA to ensure that new endpoint software releases or compression techniques won’t undermine QoE. AI VQA is capable of quantifying the QoE impact no matter where the artifacts are introduced.
View Pixels Like a Human
The new standard of video QoE analysis is a non-reference algorithm that can look at the video just like a human would and then give it a VMOS that tightly correlates with what a human would see. As its name implies, a non-reference algorithm doesn’t require access to the source video to make a comparison, thus making it practical for real-world use. Non- reference video testing is critical for live content because, unlike on-demand, you don’t have a source available that you can use for comparison.
For example, a network provider could use this approach to evaluate the QoE
SEPTEMBER 2025 Volume 47 No.3
103
Made with FlippingBook - Online magazine maker